flan/BUILT.md

# How the parts that exist work

The reasons behind the code, kept separate from `NEXT.md` so that what to do next is not buried under what was already
done. Nothing here is a plan. Everything here is load-bearing at least once: why nothing is ever `dlclose`d, why a call
bound at link time cannot be made to notice a redefinition, why the printer is a compile-time walk over a type rather
than a function in the runtime. Deleting it would mean deriving it again.

`NEXT.md` is the live document — what is in flight, what is queued, what is blocked, and the sharp edges. Read that
first. Come here when you need to know why something is the shape it is.

## What milestone 4 added

**`dotimes`** desugars in `check.ml` to a `Let` plus a `While` — no new IR node. The bound is evaluated once into a
hidden slot before the loop, so a body that changes it cannot change the trip count, and the loop variable is not
assignable, which makes the generated step its only writer.

**`defer`** is recognised in `check_fn` and nowhere else, because that is the only place that knows a form is at the top
level of a function body. Each one is checked in place, then registered on the context; it emits nothing where it
stands. Function exit runs them innermost-first, and an explicit `return` runs the ones registered *above* it — a defer
written below a return has not executed yet and must not fire. A trap runs none of them, which follows from the
bounds-check shape (`noreturn` then `unreachable`) rather than being a separate decision.

`defer` inside a `let`, a loop or a branch is **rejected**, not accepted with function scope. It would run once at
function exit rather than once per iteration, and that is the silent-wrongness class the rule below is about.
Block-scoped defer is real work and is not done.

**New builtins:** `zeroed` (takes its type from the place it is stored into), `min`/`max` (each operand through a slot,
so neither is evaluated twice), `bit-and`/`bit-or`/`bit-xor`/`<<`/`>>` (integers only; `>>` is arithmetic on a signed
type and logical on an unsigned one), and `rand-f32`.

**`rand-f32` is in the prelude, in Flan** — PCG-XSH-RR 32 over a `u64` state. It is not libc's, because a grid hash is
only a regression test if the sequence is byte-identical on native and wasm32 (plan.org, RNG is ours). `rand-seed` sets
the state. This is what the bitwise operators were added for.

**Enums and keywords.** `(defenum Name [member value ...])` gives a type that is an `i32` at run time and its own type
in the checker, so `:space` at a call site resolves against the parameter's enum and a typo is an error there rather
than a wrong number later. A keyword means nothing where no enum is expected — there is no keyword type to fall back on.

## Why the FFI goes through a C shim

The decision that shapes the whole raylib package. What clang generates for raylib's own prototypes on x86-64:

```
Vector2 {float,float}   →  declare <2 x float> @GetMousePosition()
Color   {u8,u8,u8,u8}   →  declare void @ClearBackground(i32)
Rectangle {4 × int}     →  declare { i64, i64 } @mkrect()
```

None of those is the struct's own LLVM type. A small aggregate's calling convention is not part of its layout — it is a
per-target classification the *caller* has to reproduce, and x86-64, arm64 and wasm32 classify differently. Putting that
in `emit.ml` is three classifiers to write and then keep correct forever, and a mistake shows up as `(.y m)` returning
garbage rather than as a link error.

So the boundary has one wrapper per binding, each one flattening the aggregates: a struct returns through an
out-pointer, a struct argument is passed by pointer, a Flan string crosses as ptr+len and the shim NUL-terminates a
copy. clang classifies all of it, per target, for free. `check.ml` enforces the rule — an aggregate in a `declare`
signature is rejected with the reason — so the boundary cannot quietly acquire one. This is plan.org's "one narrow host
ABI, implemented twice", and `flan_rt.c` is the same pattern.

### The wrappers are generated now — `declare-c`, `lib/shim.ml`

The price above was a hand-written wrapper per raylib call, and the prediction that they were mechanical enough to
generate "if that ever becomes the bottleneck" came true at 84 of them. `vendor/raylib/shim.c` is gone; the directory
holds `raylib.flan` and `link` and no C at all.

One binding is now one line:

```
(declare-c draw-texture [t Texture2D  x i32  y i32  tint Color] "DrawTexture")
```

`declare-c` names raylib's own function in raylib's own signature, and the compiler emits, into a C file compiled like
any other: the typedefs for the structs involved, made from the Flan `defstruct`s; the `extern` prototype in the
function's true signature; the wrapper that flattens it; and the flattened `declare` the Flan side calls, with an
ordinary Flan `defn` above it when the signature has a struct in it. `flan shim <file>` prints the whole file.

**It is a second form and not a change to `declare`, for one reason worth remembering:** `(declare start-raw [path
string] i32 "flan_agent_start")` in `vendor/agent` means the symbol takes ptr+len, and `(declare-c init-window [w i32 h
i32 title string] "InitWindow")` means it takes a NUL-terminated `char *`. Same shape, opposite claims, so no structural
rule can separate them. `declare` is untouched, and `sqrtf` and the agent still work unedited.

**What the generator guarantees, and what it trusts.** Guaranteed: the C typedef and the Flan struct come from the same
`defstruct`, so they cannot disagree — permute the `defstruct` and the typedef permutes with it, which is exactly what
makes the permutation runs below meaningful. And clang type-checks the wrapper against the generated prototype. Trusted:
that the `defstruct` matches the library's real struct, and that the `declare-c` signature is the function's real
signature — no header is read, deliberately, so nothing can check either. A `_Static_assert` on `sizeof`/`offsetof` was
considered and rejected as circular: both sides would come from the same field list. Padding is not a separate hazard:
for every field type the generator admits — the machine integers, the two floats, `bool`, a pointer and a nested struct
— LLVM's layout is C's, and `emit.ml` writes no datalayout, so clang applies the target's rules to both halves.
Everything where they could diverge is refused at the field.

**One thing got sharper and should be said plainly:** the prototype is now generated *from the declaration*, so a
scalar's width carries ABI weight it did not before. `f64` where raylib says `float` used to be narrowed by clang at the
hand-written call site; now it emits `double` and raylib reads garbage. All 84 migrated prototypes were diffed against
the deleted `shim.c`'s — which was the ground truth for the true signatures — and agree.

**Strings.** The hand-written wrappers sized the NUL-copy per call site: 256 for a window title, `PATH_MAX` for a path,
512 for drawn text, truncating past it. A generator has no call site to look at, so it must not be the thing deciding a
string is too long: 256 bytes on the stack, the heap past that, freed after the call. The only truncation left is on
malloc failure, where the alternative is handing C a null pointer.

**Two bindings keep a hand-written wrapper, and both wrappers are Flan, not C.** `collision-point-poly?` takes a slice
and `collision-lines` answers with an `Option`; neither is raylib's own signature. A slice parameter in a `declare-c` is
refused by name, because a slice's length crosses as i64 and the type of the C count parameter beside the pointer is not
recoverable from `[T]` — so that one declares `(Ptr Vector2)` with an explicit `count i32` and the Flan wrapper passes
`(addr (at points 0))` and `(len points)`. Every other refusal — an Option, a union, a fixed array, a map, a returned
string, a callback, an unknown type, an unrepresentable struct field, two Flan names for one C symbol — is by name with
the reason, and the acceptance table asserts on the reasons.

**Known edge, not fixed:** a REPL redefinition that introduces a *new* `declare-c` cannot work. `Build.shared` is llc +
`ld -shared` and compiles no C, so the wrapper would not exist in the running process. Editing the body of a function
that calls an existing binding is unaffected.

`raylib.flan` carries the nice signature and the compiler writes the rest, so the surface sand.flan sees is
`(rl/get-mouse-position)` returning a `Vector2`. Verified end to end, headless: `GetColor(0x11223344)` comes back as `17
34 51 68`, four separate bytes — a `Color` is *not* the little-endian reading of the packed integer, so an identity
would have passed a weaker test. That case is in the acceptance table, skipped if `libraylib` is not installed.

The bindings are 171 calls across thirteen structs: window, keyboard and mouse; drawing (rectangles, circles, lines, triangles, rings, ellipses, text); the eleven `collision-*` predicates; textures; the Image family; `Camera2D`; `RenderTexture2D`; the whole audio surface (device, `Wave`, `Sound`, `Music`); fonts and glyphs; and gamepads, touch and gestures — plus the `Key`, `MouseButton`, `TraceLogLevel`, `GamepadButton`, `GamepadAxis` and `Gesture` enums, raylib's own named colour palette, and the `FLAG_` window hints. Adding one is a single `declare-c` line; there is no C to write.

Two things the ported examples in `examples/` wanted and could not have, both refused for reasons that are right. `GetGamepadName` returns a `char *` into raylib's static storage: *the return type of get-gamepad-name is a string, and a string only crosses as a parameter — a C function that* returns *one returns something Flan has no owner for*. And an enum parameter cannot be indexed — `GetGamepadAxisMovement` takes a `GamepadAxis`, a loop variable is an `i32`, *expected rl/GamepadAxis, found i32*, and a second `declare-c` of the same symbol with an `i32` face is refused too: *one declare-c per C function, and another Flan name for it is a defn* — which cannot help, because a wrapper renames and does not retype. The caller spells the loop as a `cond` over the members it knows.

The texture calls are the first ones with no headless test, because loading one needs a GL context. What the acceptance
case does instead is pin the two new struct layouts using the only things raylib computes from those fields without a
GPU: `GetCollisionRec`, which pins `Rectangle` completely, and `SetShapesTexture`'s default substitution, which pins
`Texture2D`'s `id` and `format` and nothing else. Handing a struct over and reading it back proves nothing at all —
storing and returning is symmetric, so a permuted layout comes back permuted the same way and the case passes. `width`,
`height` and `mipmaps` are therefore checked only by looking at `sand.flan` running, which draws the brush sprite four
ways for that reason.

No raylib headers are needed: the generated C declares the prototypes it uses, so the build depends on the shared
library being linkable and not on `raylib-devel`. `vendor/raylib/link` carries `-l:libraylib.so.550` because Fedora
ships the runtime library without the `.so` symlink.

### The header is read now — `headers`, `lib/cimport.ml`

The section above ends by naming what the generator *trusts*: that the
`defstruct` matches the library's real struct, and that the `declare-c`
signature is the function's real signature. "No header is read, deliberately,
so nothing can check either." A header is read now, and both are checked.

**The dependency, which is the crux, and which this project already answered
once.** Zig's old `@cImport` ran clang as a *library*. That is exactly the
dependency plan.org rejected in "Why LLVM IR as text": a version-pinned C++
library breaks routinely on upgrade, a binary on PATH does not. So this shells
out for `clang -Xclang -ast-dump=json -fsyntax-only`, which is the same binary
every build already runs and adds nothing that is not already being paid for.
`lib/cjson.ml` is enough JSON to read that dump and no more, so it adds no opam
package either.

A note worth recording, because it strengthens the argument rather than
weakening it: **Zig has since abandoned clang here too.** `translate_c.zig` is
gone; `lib/compiler/translate-c/` is built on Aro, a C frontend written in Zig.
Their reason was to ship a compiler containing no clang at all — the opposite
premise to this one, where `clang` on PATH *is* the toolchain assumption. Both
projects walked away from linking libclang; only the destination differs.

**What is imported: functions, and only functions.** Not structs, not enums,
not macros. The bound on how much is not a curated list but the package's own
`defstruct`s — a function whose signature mentions a struct the package has not
described is refused with that reason, so `vendor/raylib` describing thirteen
structs is what makes the import thirteen structs wide, and describing a
fourteenth widens it. Of raylib 5.5's 581 functions, 256 import, 153 are
refused, and 172 are left alone because the package already binds them by hand.

Not generating `defstruct`s is what makes the check possible at all. Generate
them and the header becomes the authority on layout, and comparing the
package's layouts against the header's would be comparing the header with
itself. A `_Static_assert` on `sizeof`/`offsetof` was rejected in the section
above as circular for exactly that reason; **this is not circular, because the
two sides have different authors.** It is the cheapest real closure of the
trusted-not-guaranteed gap.

**Refusing by demotion, which is the one thing taken wholesale from Zig.** Zig's
translator never drops a declaration it cannot handle: `failDecl` binds the name
to a `@compileError` carrying the reason, so the name still exists, the program
still compiles, and asking for *that one name* fails at the use site with the
reason. A wholesale import has a hundred and fifty refusals and a caller cares
about the one they typed. Flan already had that mechanism — `Load.refuse_hidden`,
built for `main` — so `rl/get-gamepad-name` is not a name, and a program that
writes it is told *the return type is a string, and a string only crosses as a
parameter* rather than "unknown name".

That is also the split `shim.ml` needed and did not have. It refuses through
`Loc.fail`, which is right when a human named one function and wrong for a
wholesale import, where one returned `const char *` would kill the header. Same
judgement, different disposition; a hand-written `declare-c` still hard-fails
and `shim.ml` is untouched.

**Two C spellings mean something in a parameter that they mean nowhere else.**
`const char *` is a string going in, and the generator already knows how to hand
one over. `char *` without the const is very often a buffer the callee *writes*,
and handing it a NUL-terminated temporary would lose the writes with no
diagnostic anywhere — const is the only thing in the header that separates the
two, so it is what decides, and an out-buffer keeps a hand-written binding
saying `(Ptr u8)`. `long`, `size_t` and the rest are refused rather than
guessed, and for a reason specific to this project: it builds for x86-64, for
wasm32-wasi and for the browser, and `long` is 64 bits on the first and 32 on
the others, so a guess would be right for the target that gets tested and
silently wrong for two that do not.

**Naming.** `rl/InitWindow` is `rl/init-window`. Reversibility is not a property
of the rule — the C symbol is stored verbatim in the declaration, so the wrapper
reads the library's spelling rather than reconstructing it. What the rule must
be is *injective over one header*, since two C functions arriving under one Flan
name would surface as a duplicate declaration about a name nobody wrote. A
boundary goes before an uppercase letter after a lowercase one, before an
uppercase letter between an uppercase and a lowercase, and before a digit after
a lowercase; nowhere else. So `SetTargetFPS` is `set-target-fps` and not
`set-target-f-p-s`, `BeginMode2D` is `begin-mode-2d`, `UnloadUTF8` is
`unload-utf8`. raylib's 581 names are injective under it. **When two do collide,
neither takes the name** — resolving by order would mean that moving two lines
in somebody else's header silently rebinds a name a program is already calling.
Both are refused, both say why, and the author binds the one they want with a
`declare-c`.

**Where it runs.** `headers` beside `link` in the package directory, read the
same way: a path, any clang flags it needs, `${NAME}` expanded from the
environment. What comes back is ordinary `declare-c` declarations, generated
before the package's names are qualified, so they arrive as `rl/…` exactly like
the hand-written ones and nothing downstream can tell which is which. No new
form, no new `decl_kind`, no reader or parser change. A C symbol the package
already binds by hand is left alone, so `declare-c` stays the escape hatch and
stays the thing that wins.

A leading `?` makes a line optional, and `vendor/raylib` uses it. "No raylib
headers are needed" is a real property — a build needs libraylib linkable, not
raylib-devel installed — and requiring a header would take it from everyone in
order to give the check to whoever has one. Unset `FLAN_RAYLIB_H` and the build
is exactly what it was; set it and every signature is checked. A path that is
*set and wrong* is an error naming it, because silently behaving as though
nobody had opted in is the difference between an opt-in and a trap.

#### What the diff found

The evidence the whole lane exists for. Against raylib **5.5** — the version
whose `.so` `link` names — **all 16 `defstruct`s and all 172 hand-written
`declare-c` agree exactly.** The half BUILT.md called trusted is now checked,
and it was right.

That is only worth stating because the check has teeth. Against the **5.1-dev**
header installed in `/usr/local` it reports ten differences: nine functions that
version does not have (`CheckCollisionCircleLine`, the six `Is*Valid` renames,
`DrawRectangleRoundedLinesEx`) and `DrawRectangleRoundedLines`, which gained a
parameter. Picking the wrong header is therefore loud, which matters, because
the two headers are on the same machine and only one matches the linked library.

Both comparisons run **at build time** and stop the build, not just in the tool.
Verified by breaking them: a permuted `Texture2D` fails naming the field that
moved, and `f64` where raylib says `float` fails naming the parameter — which is
the hazard the section above calls out by name and says only a test can catch.
The message points at the line in `raylib.flan`, not at the header.

`flan import-c <header> [package.flan…]` prints what it would produce, what it
refused and why, and both comparisons, without building anything. That also
makes "generate once and commit the result" available for the cost of a
printer — explicit in the source, checked against reality, no header read at
build time.

#### What it costs, measured

The number that decides how much to import, because `reach.ml` was the reason to
think a wholesale import could be free.

| | today (172 by hand) | + 256 imported |
|---|---|---|
| release build, cold | 0.298s | 0.312s |
| release build, warm | 0.078s | 0.082s |
| redefinition (`flan reload`) | 31.0ms | 46.5ms |
| dev build, cold | 0.649s | 0.982s |

**`Reach.link` already drops a generated wrapper whose declaration nothing
reachable calls, and that is what makes the release column nearly flat.**
Confirmed on the case it exists for: a wasm32-wasi build of a program that
imports raylib and calls none of it still links without libraylib, with 256
extra declarations in play. Dev builds are not pruned, on purpose, so one
compiles all 428 wrappers — once, at session start, since `Build.shared` is
llc + `ld -shared` and compiles no C.

Reading the header is cached, and the cache earned itself against a measurement
rather than a guess: 64ms of a 72ms check, against 8ms for the whole program
without it. What is cached is the *extracted* signatures and not clang's JSON,
because the parse is half the cost — 30ms is clang writing 1.8 MB and the rest
is reading it. Keyed the way the object cache is keyed, on everything that could
change the answer: the header's path, size and mtime, the full flag list, and a
format version, since the value is marshalled. That takes the delta to 17ms.

**The 15.5ms on redefinition is the real cost and it is the argument against
importing at build time**, on the branch where the dev loop is the priority. It
is the strongest case for the third option — generate from the header, commit
the result, regenerate when the library moves — and that decision is open.

#### Where that 15.5ms actually is — and it is not the header

Written above as though a redefinition were re-reading the header on every
reload. It is not, and the number was never taken apart. Taken apart now, with a
timer around each stage of the same `flan reload` (raylib 5.5, sand.flan, one
`defn` in the forms file, warm caches):

| | no header | header imported | delta |
|---|---|---|---|
| `Session.create` | 7.0ms | 21.5ms | +14.5 |
| `Session.eval` | 4.8ms | 8.4ms | +3.6 |
| `llc` + `ld` | 21ms | 21ms | 0 |

And inside that `create`, the header's own share: reading the cached dump
**0.33ms**, turning it into declarations **3.3ms**, checking the package's
layouts and its hand-written signatures against it **0.55ms**. Roughly 4ms of
the 14.5, once. The other ~10ms is `Load` and `Check` doing their ordinary work
over 256 more declarations, and so is the +3.6ms on `eval`.

**So a `C-c C-c` in a daemon pays no header cost at all, and never did.**
`Session.eval` puts the evaluated forms through `Load.program`, and a form list
with no `(import …)` in it reads no package and therefore no header. The import
runs when the session is created, and again on a `C-c C-k` whose buffer carries
the package's own `import` line. `flan reload` is a fresh process per
redefinition, which is why its number carries a session's startup cost — it is
the CLI's shape, not the dev loop's.

What a redefinition does pay for an imported package is the +3.6ms of checking
and emitting against a program with 256 more declarations in it. That is in
`Check` and in `Emit.redefinition` declaring every sibling; it is real, it is
the number to attack next, and no cache touches it.

#### Two caches, and what a header change mid-session means

Both levels exist now. **On disk**, keyed on the header's realpath, size, mtime,
the full flag list and a format version — the object cache's own convention, and
it now lives in the object cache directory rather than a second one beside it.
Cold that read is 50ms of clang and parse; served from the file it is 0.33ms.

**In the session**, two tables in `Cimport`: the extracted dump by header, and
the generated declarations by header *and* by what the package already declares
(names taken, structs and enums known, C symbols bound by hand), so a package an
evaluation has added a decl to is worked out again rather than served a stale
answer. A repeat import in one process goes from 3.65ms to nothing — measured on
a whole-buffer evaluation, which imports twice.

**A header edited mid-session does not take effect until the session is
restarted**, and that is deliberate rather than incidental: the in-memory key is
the header's path and flags, with no mtime and no `stat`. It is the rule a
changed `.c` file already follows — the daemon compiled the package's C at
startup and a redefinition does not recompile it — and the rule the running
program follows, since its struct layouts are the ones it was built with. Taking
a new header mid-session would mean type-checking new signatures against a
process still running the old layouts, which is the silent disagreement the
whole header check exists to catch. The disk cache does key on mtime, so the
next session reads the new header rather than the old dump.

One attribution to not repeat: the disk cache is **not** what the dev build's
+333ms cold is about. Measured on this machine, a cold dev build of sand.flan is
1.37s without the header and 1.83s with it and both caches cold; clear only the
header cache and the same build is 0.92s against 0.83s. So the header read is
~60–90ms of a fresh session and the rest of the cold delta is the object cache
compiling a shim with 428 wrappers in it, which is the object cache's business
and already warm after one build.

### What a headless FFI test can and cannot pin

Worth knowing before writing another one, because two plausible tests in a row turned out to check nothing.

- **A struct round trip is worthless.** Hand raylib a struct, read it back, compare: store-and-return is symmetric, so C
writes and reads the same wrong slots and the test passes for *any* field order. Found by permuting two fields and
getting identical output.
- **Axis-aligned geometry cannot pin `Vector2`.** Exchanging `x` and `y` is a reflection, applied to the inputs on the
way in and undone on the way out, so the printed answer is unchanged. Every collision predicate, and every distance,
passes with the fields swapped — verified by swapping the `defstruct`, which is what the typedef is now made from.
Distances are worse: the reflection does not even reach them.
- **What does pin `Vector2` is the rotated camera**, because a 90-degree rotation is not axis-aligned and therefore does
not commute with the reflection. That case is load-bearing and must not be deleted on the grounds that the collision
cases look like they cover it.
- **What pins `Rectangle` is arithmetic on its fields** — `GetCollisionRec` computes four numbers from four different
field pairs, and the point/rect predicates turn the wrong way when width and height are exchanged.

- **Scalars in, fields out is the strongest shape there is**, and the Image family is where it was finally available.
`GenImageColor(4, 2, colour)` is handed two integers and answers with a struct reading 4, 2, 1, 7 — four distinct values
in four adjacent `i32` slots, with no input struct for a permutation to cancel against. That pins `Image` completely,
including that `data` is present and first; `Texture2D` could never be pinned that way because nothing without a GPU
reads its width, height or mipmaps at all.
- **A non-square image is an axis discriminator.** `GetImageColor` indexes `y*width + x`, so on a 4-wide, 2-tall image
`(3,0)` exists and its transpose does not: exchange the wrapper's `x` and `y` and the read goes out of bounds and
answers transparent black. `ImageFlipHorizontal` against `ImageFlipVertical` says the same thing twice more.
- **A file is external ground truth**, so `ExportImage` then `LoadImage` is not the symmetric round trip the rest of the
package has to avoid — stb's encoder and decoder agree with each other, not with Flan's field order. Verified red by the
`x`/`y` permutation above.

The rule that falls out: make raylib **compute** something whose answer differs per axis, then verify the test can fail
by permuting the fields and watching it go red. A case not verified that way is decoration.

One correction to an assumption that has now cost two lanes a guess: **`MeasureText` is not headless material.** It
measures with the default font, `LoadFontDefault` is not exported, and nothing but `InitWindow` loads it — so with no
window it answers 0 for every string. Measured against `libraylib.so.550`, not reasoned about. `GetFrameTime`, `GetTime`
and `GetScreenWidth`/`Height` are all 0 headless for the same kind of reason. All five are bound, and all five are
exercised by running `sand.flan` and looking, which is the whole of what can be claimed for them.

## Packages

`lib/load.ml` resolves `(import rl "vendor:raylib")` before the checker runs. The directory is the package; `vendor:` is
a collection, resolved by walking up from the importing file until a directory of that name is found; a path with no
collection is relative to the importing file. Importing is a **rename**: every top-level name the package declares
becomes `alias/name`, and every use of one — in a type, in a body, in a struct literal, in an *array length* — is
rewritten to match. Local bindings shadow. Nothing downstream knows a package existed; the checker sees one flat list of
declarations whose names contain a slash.

A package may also carry the C it binds to: every `.c` file in the directory is compiled into the build, and a file
named `link` lists extra linker arguments. Whether those reach the build at all is decided *after* checking — see below.

**A package may be a single `.flan` file** named outright, rather than a directory. That is for the program that is also
a library: `sand.flan` shares the repository root with three other loose programs, so naming its directory would import
all four. A file carries no `.c` and no `link` file; those belong to a directory.

**A package may import a package.** The qualification flattens to the *inner* alias — raylib imported by a package that
is itself imported is still `rl/…`, never `sand/rl/…` — because a directory reached along two routes has to arrive under
one set of names or the checker sees every declaration twice. It is the one decision here that would be hardest to
change later, because a qualified name is an identity the `layout` op and the break loop both resolve by.

A directory is keyed by its real path and read once, so a **diamond** shares one copy of its bottom package.
`programs/pkg-diamond.flan` is the case that proves it at the level that matters: `area/` and `draw/` both import
`shape/`, a `shape/Box` is built inside `area/` and passed to a function declared inside `draw/`. Read `shape/` twice
and there are two structs both named `shape/Box` which do not unify, so what proves the dedupe is that the program
prints `3 6 20` and not that it compiles.

The same directory under two *different* aliases is refused, including when one of the two is a package's own import
and the other is pages away in the entry file — `programs/pkg-alias-clash.flan`.

**A ring is refused and named.** Earlier it was not: the read-once table swallowed the second arrival, so mutually
dependent packages appeared to work, and BUILT.md said so. What that cost is a definite package order, which is what
the macro expander needs — every `defmacro` has to be compiled before anything that calls it — so it is now a refusal
that names the ring, `a -> b -> c -> a`, rather than the single import that happened to close it. The two questions are
kept apart by two pieces of state: the chain currently being read, and the set already finished. Found in the first is
a cycle; found only in the second is the diamond's second route and still a no-op.

`Load.t.pkgs` comes back **dependencies first** — the topological order the acyclic rule buys. The declaration list is
deliberately unsorted: `check.ml` collects every top-level name in one pass before it checks any body, so top-level
names are order-independent by construction. The order exists for the expander, which cannot work that way.

**Visibility is one rule: `main` is not exported.** A package carrying one would collide with the importer's the moment
anything imported it, so a program could never be a package; and `main` is a reachability root, so an imported one would
keep everything it calls alive. Writing `sand/main` is refused at the line that wrote it, with the reason — left to the
checker it would be "unknown name", which is true and useless.

Still missing: a package-private marker for anything other than `main`, which is why `rl/get-color-raw` is callable.
The gap is surface syntax and not `load.ml` — `exported` is one predicate and the refusal machinery that points at the
line which tried already exists, so a second rule is a line. What does not exist is any way for a package to *mark* a
name private, and inventing one is a reader and parser change.

## The link follows the program

`Load` used to hand a package's `.c` files and `link` arguments to the build the moment it was imported, whatever the
importing program did with them. So anything naming `vendor:raylib` linked libraylib on every target, and on wasm32 that
link cannot succeed.

`lib/reach.ml` answers it from the checked program instead. Start at `main` and at the globals that run before it,
follow every call — including the `Handled` frames, where a lifted handler clause is reached by address and by nothing
else — and keep what is reached. A package none of whose externs survive contributes no C and no linker argument.

Dropping the flags alone would only move the failure: the bodies that called into raylib would still be emitted and
`wasm-ld` would fail on the symbols rather than on the argument list. So the same walk prunes **functions and externs**
from the program. Only those. Globals, structs and unions stay, because an unreferenced global is bytes in BSS and a
dropped one is a silently different program.

**Dev builds are not pruned.** What a REPL may redefine next is not a function of what has been called so far.

The filtering happens at the call sites — `bin/main.ml`, the tests — because `Build.executable` receives `csrcs` and
`lflags` from its caller and never sees the import list. `Reach.link` returns the pruned program and its C and linker
arguments together, so a caller cannot take one without the other.

## sand.flan is one program

It was two files, and only ever for the reason above: the headless run is the one CI does on native *and* wasm32, and a
program that imported raylib linked libraylib whatever its `main` did. So the simulation lived in `sand-sim/` and both
drivers imported it.

Now `sand.flan` holds the simulation *and* the raylib front-end, and `test/programs/sand-headless.flan` imports
`sand.flan` itself — window, raylib bindings, dev agent and all — and still builds for wasm32. Nothing it calls reaches
raylib; `sand.flan`'s `main` is not exported, so the only `main` is the headless one; and the hash is unchanged on both
targets at `-O2` and `-O0`, which is the point. A refactor that moved that number would have moved the simulation.

What still justifies *two entry points* is smaller and stands on its own: **the headless test needs no window and no
input on any target.** `sand.flan` could not be that test even natively — with no mouse the grid stays empty and
`settle` and `move-grain` never run on real data. Measured through the probe: 168 grains painted around row 4–8, still
168 after 40 frames, lowest occupied row 68. Grains fall, and none are lost.

Two claims that got run together in an earlier note, for the record:

- *raylib does not work on wasm* — false. It works through emscripten. What is true is that it does not work on the
**wasi** path, which is what the headless table targets, and which has no GL and no browser.
- *a game loop cannot be expressed on wasm* — false. The browser cannot be blocked, so a web build drives the loop with
`emscripten_set_main_loop` instead of a `while`. ~~That is a different `main`, not a different program.~~ **The premise
held and the conclusion did not.** It is the same `main` and the same program: `-sASYNCIFY` answers the same browser
fact without cutting anything in half. See "The browser is the third target" below.

**Three edits were made to sand.flan's own text** when it was ported, and they are language decisions rather than fixes:

- `(defconst gravity 0.05)` → `(defconst gravity f32 0.05)`. An untyped float constant is `f64`, `velocity` is `[f32]`,
and there is no implicit widening.
- `(defvar current-color u32)` → `i32`. It is an index into `colors`, and `(len colors)` is an `i32`.
- `(defn main [])` is unchanged — the short form, as plan.org says.

Painting is on **hold left mouse button** rather than on space, since the mouse bindings exist now. Space is still what
cycles the colour, on release, which is a leftover and probably wants to move to the right button or to a key press.

## Bounds checks — done at milestone 3

`at` and `slice` emit `icmp` → `br` → cold block → `call` → `unreachable`; a failure names the source location. Three
check sites: `at` on `[n T]` (static bound, folded by LLVM for a literal index — and a literal that is out of bounds
never reaches emit, `check.ml` rejects it), `at` on a slice or string (runtime len), and `slice` (two comparisons — `lo
<= hi` is not redundant, without it a reversed range yields a huge unsigned length). All comparisons unsigned.
`Build.opts.checks` is on by default and **not** tied to `opts.opt`, which is what lets the acceptance table run the
same programs at `-O0` and `-O2` with identical checks. The flag is `--no-bounds-checks`.

The write path is its own case: `(set (at arr n) …)` lowers through `place`/`Pindex`, not through `At`, so a refactor
that split them would break the write check silently. The test covers both.

Cost, measured: a 50M-iteration dependency chain over a 1024-element array runs at 0.11–0.12s checked against 0.12–0.13s
unchecked. Indistinguishable.

## Sanitizers, and why hand-written IR does not get them for free

`--sanitize` builds the whole program under ASan and UBSan — the runtime's C, the generated shim, and the Flan. The last
of those is not what passing `-fsanitize=address` to the clang run over the `.ll` gets you, and the gap is silent.

**AddressSanitizer is an LLVM pass, but it instruments only functions carrying the `sanitize_address` attribute.** That
attribute is put there by clang's C frontend. `Emit` writes `.ll` by hand, so it wrote none, so the pass walked past
every Flan function and instrumented `flan_rt.c`. The measurement that settled it: an out-of-bounds read of a `defvar`
array in a `--no-bounds-checks` build printed its garbage and exited 0; with an `attributes #0 = { sanitize_address }`
group named on every `define`, the same program reports `global-buffer-overflow in flan.main`. Globals are the exception
— the module pass redzones them whether or not any function is attributed — which is why the *shape* of a sanitized
build looked right long before it worked.

**UndefinedBehaviorSanitizer has no equivalent lever.** Its checks are not a pass: the C frontend emits branches to
`__ubsan_handle_*` inline, and no attribute asks anything to produce them. So UBSan covers the C and nothing else, and
`(<< 1 32)` is still unremarked under `-fsanitize=undefined`. Shift UB, alignment and the f32→i32 cast on NaN are
therefore a compiler feature if they are wanted — checks emitted from `Emit` behind the flag, the same shape the bounds
checks already have — and not a flag away. `test_sanitize` pins both halves with controls: one program that must report
and one that must not, so either fact changing is a test failure rather than a discovery.

`-fno-sanitize=signed-integer-overflow` is the only exclusion, because wrapping is what this language's arithmetic
means and without it every program trips on its first `+`.

**The flag deliberately does not force `-O0`,** unlike `--debug`, whose reason (mem2reg deletes the alloca a
`llvm.dbg.declare` describes) does not apply. The optimiser is half of what is being measured, and `bounds.flan` proves
it: with checks off, its read past the end of a string constant is reported at `-O0` and silent at `-O2`, because an
out-of-bounds `inbounds` getelementptr into a constant is poison and LLVM folds the load away. The program then prints a
wrong answer instead of touching memory. Same family as `(<< 1 32)` compiling to a bare `retq`.

**What ASan covers of the bounds checks' job, since `--sanitize --no-bounds-checks` is the run that asks.** Three of
`bounds.flan`'s six deliberate out-of-bounds cases are caught. A negative index into a global is not, and the reason is
layout rather than anything about the access: ASan lays a global out as `{data, redzone}`, so reading before one lands
in whatever precedes it, which is a redzone if something instrumented is there and ordinary memory if nothing is.
Measured both ways — silent in `bounds.flan`, reported as soon as another `defvar` is declared in front of `arr`. A
reversed slice is not caught either, having computed a negative length and read nothing at all. And ASan sees
out-of-*object* access, not out-of-subobject, so a slice into the middle of a larger array can overrun its logical
bounds without crossing a redzone. Three of six is a ceiling on what it covers, not a measurement of the risk. It is a
second net, not a replacement.

The sweep lives on its own dune alias rather than on `dune test`: a sanitized program links to a statically linked 1.8MB
binary, and twenty-eight of them twice over is minutes against the suite's seconds.

## Why there is no interpreter

Open decision #7 is settled: **the compiled path is the only backend.** Both arguments for a permanent interpreter had
expired — the instrumentation step debugger that wanted it is cut, and compiled redefinition measured at ~16ms,
perceptually instant for expression eval too. Milestone 3 did not need an oracle either: the acceptance table is
hand-written, so the table *is* the oracle. Consequences already applied: milestone 2's "interpreted calls per second"
criterion is dropped, and the host ABI moved onto the critical path.

## The layout, which is the whole backend design

```
i8..i64 / u8..u64   i8..i64          signedness lives in the ops
f32 f64             float double
bool                i1
an enum             i32
[T] and string      { ptr, i64 }     ptr+len, non-owning
[n T]               [n x T]          inline, a value
(Ptr T)             ptr              opaque pointers
(Option T)          { i8, T }        tag 0 None, 1 Some
a struct            a literal struct, declaration order
Unit and Never      {}
```

No object headers anywhere, so a Flan struct is exactly its C struct and nothing marshals. Two consequences carry the
semantics:

- **Every slot is an `alloca`.** Reading a local is a `load`, assigning is a `store`, and a `store` of an aggregate *is*
the copy `spec-memory.md` requires. `addr` of a local is then just the alloca, and `mem2reg` removes the ones nobody
addressed. `test/programs/values.flan` pins this down.
- **A place is a pointer, a value is a load from it.** `(set (.pos c) …)` through a `(Ptr Cursor)` becomes a
`getelementptr` on the pointer, not on a copy. This is the split that would have made a tree-walker silently wrong.

Non-local exit is lowered explicitly: `return`, `some` and a failed bounds check are branches, never platform unwinding,
so wasm32 needs no exception proposal.

## The reload primitive — dev loop steps 1 and 2, measured

`llc` → `ld -shared` → `dlopen` → call, with no protocol and no daemon. `dune test` runs it: one function is recompiled
into its own object and called inside a process that is already running, twice, with a changed body the second time.

| Step | Cost |
|---|---|
| `Emit.redefinition` | below the timer (<0.1ms) |
| `llc -O2 -filetype=obj` | 15–17ms |
| `ld -shared` | 3ms |
| `dlopen` + `dlsym` | **0.04ms** |

**~19ms end to end**, and the load itself is free. plan.org's 16ms was measured with clang somewhere else; this is the
number from this codebase. For contrast, `clang -shared` on the same IR is 50ms — the driver is again most of the cost,
which is why the dev path skips it. `llc` and `clang` are both 20.1.8 here; check that before trusting the `.ll`, since
the driver absorbs IR the bare tools reject.

`ld -shared` rather than `clang -shared` for a second reason: a shared object is allowed undefined symbols, and that
*is* the mechanism. What the new module does **not** define is the whole design:

- **a global is `external`.** This settles the open question below in the only direction that supports the demo: a
redefinition can change a function's body and can never re-initialise the program's data. Define the global and the
loaded object gets a second copy — sand's `grid` would reset on every reload, and "edit the code, keep the sand" is the
thesis.
- **every other function is a `declare`**, so a redefined `settle` calls the host's `move-grain` rather than freezing a
private copy of it.
- **no `main`.** This module is loaded, not started.

Its string constants still come along; omitting them is an undefined `@.str.N` at link time, and it is easy to miss
because a one-function module usually has none. `Emit.signature` is now the single place a function's LLVM signature is
spelled, because a `define` here and a `declare` there drift the moment one of them grows a case for `Unit` or for a
slice parameter.

**`flan_dev.c` is compiled into every build, not only a dev one.** Nothing in a release build calls into it — the
compiler only emits a registry lookup for a name the host was not built with, which cannot arise without cells — but the
agent package's C refers to it, and a package's C sources are collected whatever `main` does. Leaving it out of release
builds made `flan build sand.flan` fail at the link with an undefined `flan_dev_result_read`, which reads as a compiler
bug rather than as a missing flag. The table is BSS, so the cost is address space and not binary size; `-rdynamic` and
the cells are still what `--dev` means. `test_agent.ml` links the agent program both ways for this reason.

**`-rdynamic` is load-bearing.** A normal executable exports nothing: `nm -D calc-me | grep 'flan\.'` is empty, so a
loaded module's `declare`s would have nothing to bind to. The test passes it through `lflags`, which keeps it a property
of the dev build rather than of every build. `dlsym` on `"flan.bump"` works — a dot is legal in an ELF symbol.

Two things about the test are deliberate and are what make it prove anything: both loads happen in **one process**,
since two runs would pass while saying nothing about an in-process swap; and the versions are **two paths**, since
`dlopen` caches by path and re-opening one would hand back the handle it already had, so the check would lie. And
`helper` is `(* x 2)` in one fixture and `(* x 3)` in the other: the second body is dead text, since the module declares
`helper` rather than defining it, so the expected 1024 coming back instead of 1036 is what proves the call landed on the
host's copy. With the two bodies identical nothing at run time would notice a module that grew its own.

String constants are emitted `private unnamed_addr`, so the module's own `@.str.N` cannot be interposed by the host's —
worth knowing, because with external linkage a redefined function would silently print the *old* text and nothing would
fail at link time. The fixtures each print a literal so that path is actually exercised.

### Cells — how a call site follows a redefinition

Loading a new body is not installing it. A call bound at link time cannot be made to notice one, so **a dev build routes
every Flan-to-Flan call through a cell**: a mutable global holding the address of the function that is current.

```
@"flan.cell.bump" = global ptr @"flan.bump"        ; the host defines it
%p = load ptr, ptr @"flan.cell.bump"               ; every call site
%r = call i64 %p()
```

Redefinition is then one store. A redefinition module declares the cells `external`, exactly like the globals, and
exposes `flan_reload_install()` that stores its own body into its own cell — cost **below a microsecond**, which is what
makes a frame-boundary swap a non-event.

The cell load is emitted *after* the arguments, so a redefinition landing between two calls cannot land in the middle of
one.

Four things about this that are not free choices:

- **`flan_reload_install` is a named function and not an ELF constructor.** A constructor runs during `dlopen`, on
whatever thread called it, mid-frame. The agent has to choose when the store happens. Loading and installing are
separate on purpose.
- **A redefinition's own body is `hidden`.** Default visibility in a shared object is interposable, and that applies to
*taking the address* too: plain `@"flan.bump"` inside the module resolves to the host's copy, so the installer would
publish the very function it was replacing and the reload would appear to do nothing. There is a test on the linkage,
because the failure is silent.
- **This also fixes the self-call edge**, which the previous version of this section listed as a sharp edge: a redefined
function calling itself goes through the cell like any other call, so it reaches the new body. v2 of the fixture
recurses on purpose, and would print the old body's text if it did not.
- **`-rdynamic` is what exports the cells**, so it and cells are one flag: `Build.opts.dev`, `flan build --dev`. This is
the first time `opts` means something semantic rather than an optimisation level.

LLVM cannot fold the indirection away — the cell is an external mutable global — and a `--dev` build of calc-me keeps 46
indirect calls at `-O2`. The acceptance table now runs `values`, `machine` and `sand-headless` as dev builds as well;
the sand hash is the case that matters, since it is the one result that would notice a call reaching the wrong function.

### Names that did not exist when the process started

Editing a `defvar` or a `defn` is a symbol the host exports. *Adding* one is not: there is no symbol to bind to and ELF
cannot grow one. Those go through `runtime/flan_dev.c`, which is two lookups and nothing else:

```
void **flan_dev_cell(const char *name);              /* a new function's cell */
void  *flan_dev_global(const char *name, uint64_t);  /* a new global's storage */
```

Both are idempotent, so the second module to mention a name gets what the first one got — which is the entire point. A
new global's **declared initial value travels with it**, as a constant the runtime copies on the allocation and ignores
on every call after: `calloc` alone is only right for ZII, and the "ignores afterwards" half is where "a reload must not
reset the program's state" lives. Putting it in the allocation path rather than in a branch at the call site means the
rule cannot be got wrong at one of them. The compiler picks per name: a name the host has is a symbol (one load at a
call site), a name it lacks is a registry lookup cached at install time in a module-local slot (two loads). So the
common case pays nothing for the general one.

**The unit is a list of top-level forms**, not one function — `Emit.redefinition ~fns`. `C-c C-c` passes one name, `C-c
C-k` passes a file's worth, one code path either way. It has to be: v3 of the fixture adds `extra` and uses it from a
redefined `bump`, and splitting that into two loads would leave a module referring to storage that does not exist yet.

Four rules, each of which is a silent failure if broken:

- **Every lookup resolves before any body is published.** Publish first and a caller reaches a function whose slots are
still null. Not race-testable, so it is asserted on the emitted `flan_reload_install`.
- **`flan_dev_global` refuses a size change.** The running process has already laid that memory out; handing back the
old allocation for a differently shaped type means the new body reads fields at the wrong offsets and nothing says so.
This is the layout-drift rule's first enforcement point. Retyping a var needs a restart.
- **Nothing that published anything is ever `dlclose`d.** A cell holds an address inside a module's text; unloading it
leaves every call site pointing at unmapped memory. That is a constraint on the agent too. The rule is about being
*pointed into*, which is why it has exactly two exceptions and they are not exceptions to the reasoning: a transient
thunk, which takes no registry slot and has returned; and a module the agent refuses before queueing it — no installer,
or no room in the ring — which published nothing and which nothing can name. What was leaked in the second case was the
handle *value* rather than the mapping: `dlopen` refcounts by path, so re-sending the same bad file raised a count
nothing could lower, and the one reference that could was dropped on the floor.
- **The registry never moves.** A module holds a cell's address for as long as it is loaded, so the table is fixed
capacity with a loud failure rather than growable.

The test that separates this from a plausible wrong version is **v4**, which redefines `added` — a name v3 introduced at
run time. v3's `bump` is already installed and is not rebuilt, so it picks v4 up only if its call goes through a *cell*
both modules found by the same name. Had v3 cached the function's address instead, every other assertion would still
pass and the transcript would read 246 instead of 432.

Sizes are spelled LLVM's way — `ptrtoint (ptr getelementptr (T, ptr null, i32 1) to i64)` — rather than by a layout
calculator in OCaml that would have to agree with LLVM's on every target.

### The agent — dev loop step 3

`vendor/agent/` is a package like any other: `agent.flan` declares three calls, `flan_agent.c` implements them, `link`
asks for `-lpthread`.

```
(agent/start path)   listen on a unix socket; once, at startup
(agent/poll)         install whatever has arrived; returns how many
(agent/wait ms)      the same, but waits for something first
```

The split between them is the design. `dlopen` relocates a module and takes the loader lock — milliseconds, unbounded —
so it happens on the listener thread. `flan_reload_install` is one store per function and must not land while a
redefined function is on the stack, so it happens on the game thread, at the top of the frame, when the program asks.
The two are connected by a single-producer/single-consumer ring and two atomics; the game thread never blocks on the
loader.

**A full ring is refused, at the sender, before the `dlopen`.** Of the three honest answers this is the only one that
reaches the person who asked: dropping loses a reload the sender was told was `ok`, which is the same lie more quietly,
and blocking stalls the accept loop — it serves connections inline, so a program that had stopped polling would also
stop answering `status` and `abort`, leaving the dev loop with no way to reach a program that had stopped listening to
it. The check is separate from the store because there is exactly one producer: room, once seen, cannot be taken away,
since the consumer only ever makes more of it. Sixty-four is a lot of reloads between two frames and the refusal says
what to do about it — call `agent/poll`.

**The way out of the break loop is `_exit`, not `exit`.** `exit` runs the atexit chain and the ELF destructors, which
want the loader lock the listener thread may be holding inside `dlopen`; a program asked to abort would hang instead of
dying, which is the failure the break loop exists to replace. The streams are flushed by hand at each call site, and 134
stays because that is what a trap exits with.

`wait` exists for tests. A test that races the frame rate fails on a loaded machine, so `test/programs/agent.flan` waits
for the reload instead of sleeping past it. It takes **two** reloads, which is the daemon's actual loop: the first
introduces a global the process was never built with, the second only reads it, and the second can only answer 1007 if
it found the storage the first one allocated rather than a fresh zeroed copy. One reload would not have shown that.

Two details found by running it:

- **stdout is line buffered**, set in `flan_rt_init`. The C default when stdout is a file or a pipe is a 4K block, so a
program running with a REPL attached shows nothing until it exits — and a test driving one cannot see its progress at
all, which is how this was found.
- **The reply goes out before the module is queued.** The other way round, the game thread can install and the program
can exit between the two, and the answer reaches the sender as a connection reset rather than as `ok`.
- **`ok` means queued, not installed.** The sender does not get to know when the swap happened; only the program knows
when it is between frames.

**sand.flan calls `agent/poll` at the top of its loop**, which is what step 3 was for. Verified: with sand running under
Xvfb, `flan reload sand-probe.flan game-draw` and one line on the socket, and 455 consecutive frames drew from a body
that did not exist when the process started. Building without `--dev` is fine — there are no cells, so a module is
refused on the listener thread and the loop never notices.

`flan reload <file.flan> <fn>... [-o out.so] [--new name,...]` builds one module the way the daemon will. `--new` is the
names the host was *not* built with; it is the one thing the command cannot work out for itself, and it is exactly what
the session will track automatically.

### The session

`lib/session.ml` is the program as a live thing: the declarations the running process was built from, plus every change
accepted since.

**Transactionality came for free and needed no machinery.** `Check.program` builds a fresh environment from a
declaration list on every call, so a form that fails to check mutates nothing — the accumulated list is simply not
replaced. Re-checking the whole program each evaluation costs the entire frontend, under 10ms, less than the `llc` that
follows. There is a test for the case that actually matters: a typo, then a good form, in the same session.

Two things the session knows that no single evaluation could:

- **Which names the running process was built with.** It comes from the *checked* program, not from any accumulated AST,
because `Check.program` prepends the prelude and no AST contains it. Derive it from declarations and `println` reads
as new, gets a registry cell nobody publishes, and the first call jumps to null with no diagnostic.
- **What that process's memory looks like.** Three changes are refused with a reason rather than loaded:

| Change | What it would have broken |
|---|---|
| a function's signature | a cell is a bare `ptr`; every call site compiled before the change still passes the old arguments through it — **and this is now a stopgap**, see below |
| a global's type | the storage exists and has a shape — reuse reads at the wrong offsets, replacement discards the state the reload exists to preserve |
| a struct's fields | the values the process is holding have the old layout |
| a `defconst`'s value, **when the checker consumed it** | it is in the *shape* of the program — `(defconst rows (/ h c))` decides `grid`'s type before anything else resolves — so no store can reach it |
| a `defenum` member | `:space` is erased to an `i32` literal in the caller, so it is folded there too |

**The signature row is the one the plan has moved past.** plan.org now says a signature-changing redefinition should
make a new internal function version with its own trampoline: newly compiled code resolves the name to it, while
existing callers and stored `Fn` values keep the old version and stay safe, and the session warns at every tracked
caller site still targeting the old signature — recompiling one either retargets it or gives an ordinary type error.
Open decision #6 records it the same way. None of the three parts exists: there are no function versions, no trampolines
(a cell holds a body address today), and no record of which source locations called what. So the refusal stays, because
the alternative to refusing is not the new design, it is a silent argument mismatch. It is a stopgap and the message
should not be read as the final answer.

A `defvar`'s *initial value* is deliberately **not** in that table. Its storage holds live state the program moved past
long ago, and refusing to change the initialiser would be refusing "edit the code, keep the sand". Same `Tast.global`
record as a `defconst`, opposite answers, told apart by `gconst`. The enum comparison runs over declarations rather than
the checked program, because `Tast.program` carries no enums at all — they are erased to `i32` in the checker, which is
the same fact that makes them unreloadable.

Note what the checker catches on its own: change `helper`'s parameter type and the *caller* fails to type check first,
loudly. The session's rules only get a turn on a change the checker accepts — one to a name nothing else in the program
uses, which is exactly where the silent version lives. The fixtures carry an unused `defvar` and a C-called `defn` for
that reason.

A `defconst` the checker never consumed is a different matter and **can** be changed: it is only ever bytes in memory. A
dev build emits every `defconst` as a mutable `global` rather than a `constant` — so LLVM cannot fold a read of it and a
module can store into it — and a changed one is published at the frame boundary exactly as a new function body is. That
is how sand's `colors` gets tuned live while `rows` stays refused. Release builds emit `constant` and get all the
folding back; `Tast.global.gfolded` is what tells the two apart, because nothing downstream of the checker could.

**A form typed into a file that is imported as a package is qualified the way the import qualified it.** `poll` in
`vendor/agent/agent.flan` becomes `agent/poll`, and its call to `poll-raw` becomes `agent/poll-raw` — through `Load`'s
own `qualify_decl`, so the rule cannot drift from the one used at import time. Without this the form spliced as a
brand-new unrelated name: the evaluation answered `ok`, and the running program went on calling the `sim/settle` it
already had. Since sand's simulation lives in a package, the one thing worth tuning live was the one thing that silently
did nothing.

It is derived from the file's path and **not sent by the editor**, which is where this departs from CIDER's `ns` key: a
Clojure namespace is declared in the file, but a Flan alias is chosen by whatever imported the directory and is written
nowhere the editor can see. One directory imported under two aliases is refused with the reason rather than resolved to
either.

The accumulated list is the **post-`Load`** one, so an evaluated `(import …)` is spliced as its expansion. Otherwise
re-evaluating a file that imports something appends a second import, `Load` expands it again, and the duplicate-name
pass rejects it. `C-c C-k` on sand.flan's own text is the test.

`flan reload <program.flan> <forms.flan>` is that path from the command line: a session over the program the process was
built from, and a file of the forms that changed. Verified against a running sand under Xvfb — a one-form `game-draw`
and 910 consecutive frames drew it.

Two limits of that command specifically, neither of them true of sessions: it builds a fresh session from source on
every invocation, so if the program file has been edited since the process launched, its idea of which names the host
has and what its memory looks like describes a binary that is not running. And `Session.eval`'s `origin` defaults to
`<eval>`, so an error in forms sent without one reports positions in a file that does not exist — the daemon has to pass
the real buffer path, which is the same key CIDER's `eval` carries.

### The daemon — `flan dev`

`flan dev <program.flan>` holds one `Session`, builds the program, launches it, and listens on `.flan-dev.sock` beside
the source. What it adds over `flan reload` is that the session *persists* — a `defvar` added by one evaluation is part
of what the next one is checked against — and that it **owns the build**, which is what makes its layout rules describe
the process that is actually running rather than a guess about it.

**The protocol is s-expressions, not bencode.** nREPL was the plan and the argument for it evaporated once the client
became ours too: there is no CIDER to be compatible with, `eval` is string-in/string-out with no slot for *which form,
from which file*, and Emacs already has `read` and `prin1`. So it is one sexp per message — no parsing code on the
editor side, and on this side the parser is the language's own reader, where `:op` is already a keyword and a payload of
Flan source is already a string literal. Framing is a decimal byte count and a newline, because the payload contains
newlines. An nREPL front end can sit on the same `Session` later; it should not have gated the editor.

```
(:op "describe")                        → (:status "ok" :fns (…) :globals (…) :alive t)
(:op "eval" :code "…" :file "/buf.flan") → (:status "ok" :names (…) :fns (…) :ms 19.0)
                                        → (:status "error" :message "…" :loc "/buf.flan:1:19")
(:op "defs")                            → (:status "ok" :defs ((name kind signature loc) …))
(:op "close")
```

`defs` is its own op rather than more fields on `describe`, because `describe` is what an editor *polls* — it is how the
program's output is drained — and signatures on that would be paid for every time anyone glanced at the output buffer.
It is asked once on connect and again after each accepted install. Four strings an editor reads with `read` and nothing
else: eldoc, completion and find-definition want the same three facts about a name. `loc` is empty where there is none
to give, because only `Tast.fn` carries one — an editor must refuse rather than go looking for the definition itself,
which in a program of several files finds the wrong one. Parameter *names* are not in the Tast, so a signature is `step
[i64 f32] i64`: types only.

The daemon makes its own source path absolute before building, because every location it reports derives from it. `flan
dev src/game.flan` run from a project root otherwise answered `src/game.flan:12:7`, which an editor can only resolve by
guessing which directory it was relative to.

An evaluation that declares nothing to install — a declaration the program already has, with no body and no new storage
— is accepted and answered with `:note "nothing to install"` rather than by shipping an empty module. Building one
anyway reports success for a change that cannot have taken effect, and costs the program a reload it did not need.

`:file` is not decoration: `Session.eval`'s origin defaults to `<eval>`, so without it every error an editor shows
points into a file that does not exist.

Two things the daemon must not paper over, both of which would look like a successful evaluation:

- **The agent socket is chosen by the daemon**, not by the program. A program's source has to name some path — sand.flan
says `/tmp/flan-sand.sock` — and the daemon overrides it through `FLAN_AGENT_SOCKET` before spawning. Guessing instead
fails silently: the module compiles, is built, and nobody receives it.
- **Delivery is checked.** `agent/start` returning 0 means a socket was bound, not that anyone connected. A failed
connect or a reply that is not `ok` becomes an error the editor sees.

It waits for the program to bind before accepting an evaluation — one arriving first would fail for a reason that reads
like a compiler bug — and it accepts with a timeout so that a program which has exited takes the daemon with it rather
than leaving an editor waiting on a socket nobody is serving.

### The Emacs client

`emacs/flan-mode.el` derives from `prog-mode` with `lisp-mode`'s syntax table, which is most of the work: Flan is
s-expressions, so sexp motion, paren matching, `beginning-of-defun` and indentation are already right. What it adds is
Flan's own bracket syntax (`[` and `{` are brackets, not symbol characters — every binding list and every type is
written with them), the characters a Flan name may contain (`-`, `?`, `/`, `.`), and its keywords.

`emacs/flan-dev.el` is the client. There is no parser in it, which is the point of the protocol choice: `prin1` writes a
request and `read` reads a reply.

| | |
|---|---|
| `C-c C-c` | the top-level form at point, recompiled and installed |
| `C-c C-k` | the whole buffer, as **one** module |
| `C-x C-e` | the expression before point, evaluated *in the running program* |
| `C-c C-z` / `C-c C-q` | connect (finds `.flan-dev.sock` upward) / disconnect |
| `C-c C-o` | the running program's own output, in `*flan-output*` |
| `C-c C-r` | a prompt on the running program (`*flan-repl*`) |
| `C-c C-b` | what a **stopped** program is offering, and which to take |
| `C-c C-d` | what the running program currently defines |
| `C-c C-a` | the code a name compiled to — amd64, or `C-u` for the LLVM IR |
| `M-.` / `M-,` | where a name is written, through an `xref` backend |

eldoc, `completion-at-point` and `M-.` all read one cached `defs` reply rather than asking per keystroke: eldoc fires on
an idle timer and completion inside redisplay, and neither may block on a socket or signal. The cache is refreshed at
the two moments the answer can have changed — on connect, and after an evaluation the daemon accepted — so a `defn` just
installed completes at once.

The modeline says whether there is a program on the other end, in three states. `lost` is a daemon that has gone away,
which is ordinary rather than an error — `flan dev` ends when its program does — so the next request reconnects on the
socket it was on. Strictly *before* a send, never after one: a connection that died mid-request may have died after the
daemon ran what it was given, and resending would install it twice or evaluate a side-effecting expression twice.

`C-c C-k` sends one module rather than a form at a time on purpose: a `defvar` and the function that uses it have to
arrive in the same load, or the first refers to storage that does not exist yet.

**Framing is in bytes and Emacs counts characters.** Every length goes through `string-bytes` and the process is binary,
or a single non-ASCII character in a buffer puts the reply stream out of step by exactly as many bytes as the payload
has of them — a bug that would look like a corrupt protocol and appear only for some users. `test/test_emacs.ml` drives
the real client against a real daemon for this reason: it is not the same claim as the daemon answering correctly, and a
mistake in the framing, in `beginning-of-defun` over Flan's syntax table, or in the reply reader passes `test_dev.ml`
and fails here.

An error comes back with a location and the client draws an overlay there, with the message beside the code, cleared the
next time that buffer's evaluation is accepted. Two things had to be right first. **The column in a `:loc` is a byte
offset**, because the reader walks the source a byte at a time — the same rule as the framing, in a different place, and
`forward-char` with it put the marker as many columns right as the line had non-ASCII characters before it. And the
daemon numbers lines from the start of what it was *sent*, so `C-c C-c` on a defn halfway down a buffer answered line 1
and every overlay would have sat on the file's first line; the client pads the form with leading newlines, which the
reader skips, so the reply's line numbers are the buffer's own.

An accepted evaluation says which names landed and what the build cost, and flashes the region that was sent. Silent
success is indistinguishable from silent failure, and `beginning-of-defun` may well have found a different form from the
one point looked like it was in.

**Live disassembly — done.** `C-c C-a` on a name writes the amd64 the running program's copy of it was assembled to
into `*flan-disassembly*`; `C-u C-c C-a` writes the LLVM IR that body was built from. `(:op "disassemble" :name … :form
"asm"|"ir")` is the op.

What makes it possible is that the daemon owns the build: it compiled every module it sent, so `objdump -d
--disassemble=flan.<name>` on the right object *is* the disassembly and the retained `.ll` is the IR. `Build.shared`
deletes its own `.ll` and `Build.executable` leaves the host's under a name that says nothing about which module it was,
so the daemon now writes its own copy beside each `.so` and keeps the host's as `host.ll` — ten reloads in, nothing else
on the machine still has that text. A table from function name to the last module accepted for it is the whole of the
bookkeeping.

**What it will not claim is that the code shown is installed**, and this is the interesting half. The agent's socket
takes a module path and five verbs; none of them reports an address, `flan_dev_cell` lives in the program's address
space, and `C-x C-e` renders a pointer as `<ptr>` on purpose — so nothing the daemon can ask would tell it what a cell
holds. The reply carries `:basis` saying which of three things is true, and the buffer prints it above the first
instruction:

- nothing has been delivered for this name, so the cell still holds the host's body — the one case that is *certain*;
- a module was delivered and the agent queued it, and the program installs it at its next frame boundary — unconfirmed;
- a module was delivered and the program is **stopped** — which says nothing either way about whether it installed,
since the commonest way to stop is to install a body and have it error; what is certain is only that nothing further
installs until it resumes.

From SBCL: offsets from the function's own start rather than addresses into an object, and `L0:` labels on branch
targets with the file address that duplicates them dropped. Not source interleaving — SBCL has the mapping and this
build emits no line tables — so the reply says that in words rather than printing a listing with no source in it. When
the debug build lands, that is the line to delete.

**Transient error overlays — done.** The diagnostic ghost text is feedback about the evaluation that just failed, not an
annotation on the source, so the next command in that buffer takes it down — edit, motion, evaluation, anything.
`pre-command-hook` and not `post-command-hook`, which fires at the end of the *failing* command and would clear the
overlay before redisplay had drawn it. The hook is buffer-local and lives exactly as long as an overlay does: added
where one is drawn, removed where they are cleared, so a session of twenty buffers is not running it on every keystroke
in all of them. `execute-kbd-macro` runs no `pre-command-hook` under `--batch`, so the test drives `run-hooks` — the
same call the command loop makes — and checks the hook is installed in that buffer and in no other; that Emacs runs it
is Emacs' contract and a test claiming to check it would be checking nothing.

**The program's stdout is a pipe into the daemon**, and whatever it printed since the last reply rides along with the
next one into `*flan-output*`. Having it arrive *with* a reply rather than by a separate request is the point: the
output an evaluation itself caused is the output anyone wants to see. Draining that pipe is a liveness requirement and
not a nicety — a pipe nobody reads fills at 64K and the next write blocks the program forever — so it is read from the
accept loop's `select`, not only when an editor asks, and the buffer is capped so a program printing every frame cannot
grow the daemon without limit.

### `C-x C-e` — evaluating an expression

A different primitive from redefining a name, and the difference is the whole design. There is no name to install a body
into, so the expression is wrapped in a function with nowhere to be called from; the module says *run this once* by
exporting `flan_reload_call`, and the agent calls it after the install — on the game thread, at a frame boundary, so an
expression reading the program's state sees a point the program agrees is consistent.

**Nothing is marshalled back, because nothing could be.** A Flan value carries no header, so no code at run time can say
what it is. The compiler knows the type and renders it *there*, in the thunk, into `flan_dev_result`. That is the layout
decision's bill, and it is why the printer set is small rather than universal.

It does not go through stdout. Stdout belongs to the program, it is in the hot path for anything that prints, and a
dev-only feature must not put a branch in it — so `flan_rt.c` is untouched and the value is read back over the agent's
socket. The read is safe without a handshake because the counter is a **seqlock**, and it had to be made into a real
one: the first version bumped the generation last and handed back the buffer itself, which says a new value has arrived
and says nothing about whether the bytes the agent then wrote to a socket were that value — the game thread is free to
be a hundred bytes into the next one by then. A seqlock cannot validate a read that finishes after it returns, so the
bare pointer was the bug rather than the ordering. `flan_dev_result_read` copies into the caller's buffer and checks
the counter either side of the copy; the counter is odd for exactly as long as a value is being written, and a reader
that loses the race reports the last *complete* generation and no bytes, so a daemon polling for a new value keeps
polling rather than being shown half of one. The count handed out is the number of complete values, so the daemon's
"has it moved" still means what it meant. Marking the counter odd needs a release *fence* and not a release store — a
release store orders what precedes it, so the writes to the buffer would be free to become visible ahead of it, which
is the original bug with more ceremony. And the odd mark is *set* rather than incremented, because a `begin` with no
`end` is reachable: the thunk calls `begin` before it evaluates anything, so an expression that signals is stopped
inside that window, and a restart taken from the break transfers past the thunk and `end` never runs. Incrementing
would leave the counter odd for the life of the process and every later read reporting "in progress"; setting the bit
means the next evaluation repairs it. The daemon waits for the count to move rather than assuming the program has
reached a frame boundary.

**This renderer is most of `println`**, which is worth knowing before anyone schedules it. plan.org describes a
compiler-provided, type-directed intrinsic that selects or emits a structural printer per concrete instantiation, prints
structs, fixed arrays and options structurally, prints a `Ptr` as its address rather than following it, and bounds depth
and length. That is a description of what `Session` already does for `C-x C-e` — same walk, same refusals, same three
bounds — aimed at `flan_dev_emit` and the wire instead of at stdout. What `println` needs on top is a stdout sink, a
builtin that takes its printer from the argument's type, and the `any`/`Error` dynamic cases, which have no compile-time
type to walk. Not the walk itself.

The renderer is a **compile-time walk over the type**, emitting a piece at a time through `flan_dev_emit`. Piecewise
because a struct is its fields with punctuation between them, and concatenating that in generated IR would need an
allocator the language does not have.

```
big                       18446744073709551615
col                       :blue
(.pos b)                  (V {:x 1.5 :y 0})
b                         (Blob {:id 7 :name "sandy \"quoted\"" :pos (V {:x 1.5 :y 0}) :tags [ 0 42 0]})
(slice (.tags b) 0 3)     [ 0 42 0]
(rl/get-color 0x11223344) (rl/Color {:r 17 :g 34 :b 51 :a 68})
sim/grid                  [ [ 0 0 0 0 0 0 0 0 ...] [ 0 ... ] ...]
```

Details that are decisions rather than formatting:

- **`u64` renders in C**, with `%llu`. The language's own `i64->bytes` is signed, so it used to refuse rather than come
back as `-1` — but refusing a whole struct because one field is a `u64` is much worse, so the runtime got an entry point
instead.
- **Strings are quoted and escaped**, also in C. Unescaped content does not round-trip and reads as a framing bug rather
than as the value it is.
- **An enum renders as `:name`**, recovered from the checker's table as a chain of comparisons, because members are
erased to `i32` before the backend sees them. A value outside the declared members falls through to its number, which is
exactly what you would want to see.
- **A pointer is never followed** — `<ptr>`. It is the only thing that could make the walk cycle, and dereferencing one
a REPL was handed is not a safe thing to do on someone's behalf.
- **Three separate bounds**, easy to conflate. `depth` (4) and `span` (8) bound the *walk*, so `[100 [100 u32]]` does
not become ten thousand render sites in one module. The *output* is bounded once in the runtime — `emit` truncates at 4K
and `end` appends `...` — because a slice renders through a loop the compiler cannot bound, and one place enforcing it
means no renderer carries a budget.
- A slice is the one case needing a runtime loop, and the slice goes into a slot first so the expression it came from is
not evaluated once per element.

What still refuses by name: `Map`, `Fn`, a type variable.

A caveat inherited from the language, not introduced here: `3.0` renders as `3`, indistinguishable from the integer.
`flan run calc-me.flan "1.5 * 2.0"` has always said `3`.

An evaluation is **not** a declaration: the thunk is built against the program and never spliced into it, so `describe`
does not fill up with `eval/N` for every expression ever typed.

**The module is unloaded afterwards**, which is the one case where that is safe. The thunk is called directly by
`flan_reload_call` rather than through a cell, and it takes no registry slot — so once it has returned, nothing points
into its text and the value it produced has been copied out. It declares that with `@flan_reload_transient` and the
agent `dlclose`s it. Measured: sixteen expression evaluations retain **zero** mappings, where each *redefinition*
retains three, permanently and correctly — a module that publishes a body exists precisely to leave a pointer behind,
and can never claim this.

Skipping the registry matters for more than tidiness: the table holds 4096 names and an expression evaluated in a loop
would exhaust it.

The test that matters is the same expression twice: the fixture increments `ticks` every frame, so two evaluations must
disagree. A value computed in the compiler, or read out of a copy of the program's state, would not.

### The REPL buffer

`flan-repl.el` is a `comint-mode` buffer whose every line goes through the same `eval-expr` request `C-x C-e` uses. No
new protocol and no compiler support. Deriving from `comint` rather than hand-rolling a prompt is the same call as
deriving `flan-mode` from `lisp-mode`: history, the input ring and kill/yank already exist. There is no subprocess
behind it — the "process" is a stub comint needs in order to have a prompt at all.

Three things about it that are decisions:

- **It is program-scoped.** A name typed at the prompt resolves against the running program's top-level namespace, so in
sand you write `sim/settle` and not `settle`. A buffer visiting a package's own file gets the alias applied for it
because the file says which package it belongs to; a prompt has no file and nothing to derive one from.
- **RET on a half-typed form opens a line instead of sending it.** Balance is checked with the Flan syntax table, so a
paren inside a string does not count.
- **A value and the program's output are different things and arrive by different routes.** The value is the result of
the request and appears at the prompt; anything the program printed while evaluating it rides along on the same reply
and goes to `*flan-output*`. Showing them in one place would be convenient and wrong, so there is a test for the
separation.

That test is what caught a real bug: the renderer's `Unit` case emitted `()` without evaluating the expression, so
`(println "x")` — the most ordinary thing anyone types at a prompt — answered `()` while nothing happened. A Unit
expression is almost always a call made for its effect, and is now evaluated and *then* reported.

### Conditions — step 1: `handler-bind` and `signal`

`spec-conditions.md` §1 and §2, and nothing else yet. They are worth having on their own because **neither alters
control flow**: `signal` returns `Unit` whatever it finds, a handler that returns normally leaves the signalling
function to carry on, and with nothing matching it is a no-op. So none of the transfer machinery §6 describes exists
yet, and no signature changed.

```
(handler-bind [(AssetMissing [c] (set seen (+ seen (i64 (.id c)))))]
  (load-all))
```

The runtime is a linked list: establishing a handler is two stores and a push onto a frame allocated on the establishing
function's own stack, and `signal` with an empty stack is a null check — which is what §2 asks for. Popping is by frame
rather than by count, so restoring what this one displaced is right even if something below it left the stack out of
step.

Three decisions worth keeping:

- **A condition's type is a hash of its name**, not an index. An index would shift the moment a struct were added, and
every handler a running program had already pushed would then match the wrong type. FNV-1a over the name.
- **The condition crosses as a pointer**, because a handler runs while the signalling frame is still alive and there is
nothing to copy. What the clause *binds* is the condition itself, though — the pointer is a hidden parameter and the
name is a slot loaded from it, so a handler passing `c` to something expecting the struct is not handed an address
instead.
- **A clause is lifted into a function of its own.** A handler runs from wherever the signal was, so it cannot be a
branch in the function that wrote it.
- **A pushed handler frame holds the clause's body address, not a cell.** This is a deliberate divergence from
plan.org's rule that a top-level function value is a stable trampoline over the cell and never the address of a
particular body. A handler frame is not a `Fn` value — nothing in the language can name it — and it is live only for the
duration of the `handler-bind` body, so a reload landing while it is on the stack finds the clause it pushed still
valid, which is exactly the "old code is never unloaded" guarantee. The consequence to know: a handler already on the
stack does *not* observe a redefinition of its own clause; the next entry to the `handler-bind` pushes the new one. When
`Fn` values arrive, this is the one place that stores a body address on purpose and must not be swept up with them.

Which gives the two refusals, both by the house rule rather than by accident:

- **A handler cannot see the establishing function's locals.** That is a closure with an explicit environment, so a
reference to one is refused *for that reason* rather than reported as an unknown name. Globals and the condition are in
scope, which is what the accumulation case needs.

What it needs is narrower than it looks, and worth getting right before anyone schedules it: a handler frame does not
outlive the function that established it, so this is spec-memory.md's **case 2** — a non-escaping `fn` capturing by
value into a stack environment — and *not* the escaping closure that plan.org's open decision #5 defers until a concrete
use case. Case 2 is settled, and #5 says in as many words that without it "conditions are not worth building". So the
biggest usability limit in conditions is not behind the thing that was just deferred.
- **`return` inside a `handler-bind` body is refused.** The frames are popped on the way out and an early exit would
leave them on the stack pointing into a function that has gone. Same shape as `defer` inside a block.

### The break loop — conditions step 3

Where **"a crash kills the program"** stops being true. An unhandled `error` runs a hook instead of `rt_die()`, on the
frame that erred with nothing unwound, so the condition and every restart between there and the top are still live.

```
flan: unhandled Missing — stopped, not dead.
   0. restart: retry
   1. restart: use-placeholder
```

Four decisions, each of which is the reason something is where it is:

- **It is a hook, not a call.** The loop lives in `vendor/agent/`, which is an optional package; `flan_rt.c` is the
release runtime and must not depend on something a program may never import. A program with no agent leaves the hook
null and dies exactly as it did before.
- **The hook resumes by writing a restart into the transfer channel** — the same channel an `invoke-restart` writes,
reaching the same guard. Choosing a restart from the break loop and choosing one from a handler are therefore the same
act, lowered once. Nothing about §6 needed changing to support it.
- **The break loop *is* the poll loop**, run from the error instead of from the frame boundary. That is not a
convenience: an expression evaluated while stopped is a module the listener queues and the game thread runs, so a loop
that did not drain that queue would hang `C-x C-e` exactly when it is most wanted.
- **Installing while stopped is allowed**, which contradicts a rule stated above and should. "A redefined function must
not be swapped while it is on the stack" is about *mid-frame consistency* — half a frame of old code and half of new —
and there is no frame in progress here. The old body on the stack keeps running; a `retry` restart calls through the
cell and reaches the new one. That is the fix-it-and-retry loop, and refusing the install would remove the point of
stopping.

**A restart frame carries its name now**, beside the hash. Matching never needs it — that is what the hash is for — but
a break loop has to *show* someone their choices and nothing at run time can turn a hash back into a name. It is also
`compute-restarts`' data, whenever that arrives.

**A choice is validated on the listener thread**, against a stack the stopped game thread is holding still, and refused
there. Accepting it and discovering on the game thread that no frame offers it would answer `ok` for something that
cannot happen.

The socket verbs are `restarts`, `restart-at <n> [name]`, `restart <name>` and `abort`, and all four are refused with
the reason when the program is not stopped — there is no restart stack to walk from a running one. `restarts` answers
`<index> <+|-> <name>` per line: the index is the identity, and the flag says whether a transfer to that frame has
anywhere to land. `test/programs/break.flan` errors three times — two restarts taken by name, then a shadowed pair
where only an index can reach the outer one — so neither a loop that always resumed the same way nor one that resolved
by name could pass.

### The break loop in the editor — conditions step 3, the other half

The loop above is reachable from a raw socket. This is the half that makes it reachable from Emacs, and the whole of it
follows from one fact: **a program stops at a moment nobody asked about.** Every other op in the protocol answers a
question an editor chose to ask.

So the state is learned **twice, deliberately**:

- **It rides on every reply**, beside the program's output and for the same reason. `:stopped t :condition "Missing"`,
or `:stopped nil`. The likeliest instant for a program to stop is the one just after an evaluation — a body that now
errors — and that is a reply the client is already reading. Finding out a second later from a poll would mean finding
out *after* the echo area had said the evaluation was fine.
- **And a timer asks anyway**, once a second, with `describe` — the cheap op, which is also how the output pipe is
drained. A program that stops in a frame of its own game loop produces no reply at all, and folding state into replies
that never come says nothing. The timer never *reconnects*: `flan-dev--request` reopens a socket a restarted daemon left
behind, which is right for something a person did and wrong for a background poll, because it would quietly erase the
`lost` state that exists to be seen. It also skips while another request is in flight — `accept-process-output` runs
timers, so a poll firing inside a read would eat the reply that read was waiting for.

Three ops, and the *annotation* owns `:stopped`, not the ops — one place in the daemon decides whether the program is
stopped, so the poll and the prompt cannot disagree.

```
(:op "break")                       → (:status "ok" :restarts ("retry" …) :unreachable (2 3)
                                         :stopped t :condition "Missing")
(:op "restart-at" :index 2 :name "retry") → (:status "ok" :index 2 :note "accepted; …")
(:op "restart" :name "retry")       → (:status "ok" :restart "retry" :note "accepted; …")
(:op "abort")                       → (:status "ok" :note "the program is exiting; …")
```

`:restarts` is *positional* — innermost first, duplicates kept — because the position is what `restart-at` takes.
`:unreachable` names the positions that are on the list and cannot be chosen. `restart-at`'s `:name` is optional and is
not the lookup: the program checks it against the name it holds at that index and refuses if they have drifted, so a
prompt cannot take a different restart than the one it showed.

`break` carries only the restart list, because it costs a second round trip to the program and is wanted only by
someone about to choose from it.

**`ok` from `restart` means accepted, not resumed.** The choice is validated on the program's listener thread against
the stopped stack, then taken when that thread next comes round its loop. A client that read it as "running again" would
poll once, find it still stopped, and re-open the prompt it had just answered — so the client clears its own flag and
lets the next poll settle it.

The modeline is a fourth state, `flan:stopped(Missing)`, before `live`: a stopped program looks exactly like a running
one from anywhere else in Emacs. The prompt is a `completing-read` over the names with `require-match`, which is exactly
right for a closed set the program computed, and `abort` is the last entry on that same list rather than a second key —
it is the thing you pick when none of the restarts is the answer.

**`flan_agent_poll` had to become re-entrant**, and that is the one thing here that was a bug rather than an addition. A
`C-x C-e` thunk may itself error; the break loop that catches it polls again from inside that very call. The old loop
cached `head` and `tail` and wrote `tail` back at the end, so the outer call rewound the index over everything the
nested one had consumed — and re-ran the thunk that had just stopped the program, which is an unbounded recursion of
breaks rather than a stumble. It now claims each job by advancing `tail` before running it, and re-reads both indices
each time round. Still single-consumer: only the game thread writes `tail`, nesting included. `test_dev.ml` evaluates an
expression that errors and resumes it, which fails against the old shape.

The agent grew one verb, `status`, answered in **both** states — `running` or `stopped <condition>`. Everything else the
break loop offers is refused while running, and rightly; but the question an editor asks *without already knowing* had
to have an answer either way, or there would be nothing to poll. The condition is its class name and nothing more: the
hook is handed a name and an opaque pointer, and nothing at run time can render a value whose type it does not know.

`test/programs/dev-break.flan` stops on its first frame, so the daemon meets a program that is *already* stopped — the
state an editor has to cope with and the hardest one to arrange later. The Emacs test breaks a program the other way
round, by installing a `step` that errors into a loop that calls it, fixes it while stopped, and then resumes: `C-x C-e`
answering while the program sits in the break loop is checked there against the real client, not only in OCaml.

### `layout` — a type's fields, with no program involved

```
(:op "layout" :type "sim/Cell")  → (:status "ok" :type "sim/Cell"
                                       :fields (("heat" "f32") ("next" "(Option sim/Cell)")))
                                 → (:status "error" :message "Missing is not a qualified name; …"
                                       :candidates ("a/Missing" "b/Missing"))
```

The daemon can answer this with nothing running. A layout is a fact about the *build*, and the daemon owns the build —
`Tast.structs` is sitting in the session it compiled the process from. That is why the conditions buffer can name and
type a condition's fields while every one of their *values* stays refused: the shape is knowable and the contents are
not, and drawing them apart says more than drawing neither.

**The type is a name, and the name is the qualified one.** This was the open question — a class name is not an identity,
and two packages each declaring `Missing` would leave the daemon unable to pick. It turned out to need no new
machinery: `Load.qualify_decl` rewrites `Defstruct (n, …)` to `Defstruct (alias/n, …)` at import, so by the time
anything reaches `Tast.structs` the names are one flat namespace in which a collision cannot exist. The name *is* the
type id, with no table to keep in step across a reload, and the existing spelling of a type — `Types.to_string` —
already prints it.

**And the break loop was already speaking it.** `Emit.struct_name_of` takes `Types.Named n` — the qualified name — and
passes it to `flan_error`; the agent holds it in `condition_name`; `break` answers it as `:condition`. So the string
the conditions buffer already had in hand resolves as `:type` by construction, and `test_dev.ml` round-trips exactly
that: the condition a stopped program reports, handed straight back, answers with that condition's fields. One caveat
worth writing down — `condition_name` is a `char[128]`, so the round trip holds for names up to 127 bytes and a longer
one is truncated and will not resolve.

**A bare name is refused, not resolved**, even when only one struct's last segment matches it. Resolving a unique
suffix would reintroduce the ambiguity the rule exists to remove, and a rule with an exception is one a client cannot
rely on. The refusal carries `:candidates`, so a person is one copy-paste from the answer and a client has its
completion list — the same shape as `package_of` refusing a directory imported under two aliases rather than picking
one. An enum is refused by *kind* (`X is an enum, not a struct`): its members are erased to `i32` before `Tast.program`
exists, which is the same fact that makes a `defenum` unreloadable. A union is refused the same way and for its own
reason — it is declared, and union *values* are milestone 6. Both are `Types.Named` at a use site, so falling through
to "no struct is named X" would say a type does not exist about one that plainly does.

**`render.ml` is not reused, and that is not a second walk.** It walks a *value* and emits the code that prints it;
this describes a *type* and emits text. What is shared is the spelling: field types go through `Types.to_string`, which
is what `defs` spells a signature with, so `(Option T)`, `[T]`, `[n T]` and `(Ptr T)` read the same in a layout, in a
signature and in the source. A field that is itself a struct shows its qualified name — which is a `:type` this op
accepts, so nesting is another request rather than a recursion, and nothing here can be made to walk forever. Prelude
structs are answered like any other, because `Render` resolves against the same list and an editor that could see a
`Split` printed but not ask about it would be the two disagreeing.

On the Emacs side `flan-cnr-layout` makes the request and `flan-cnr-show` passes the result into
`flan-cnr-state-from-reply`, which stays a function from data to data so the fixture-driven tests keep working without
a socket. A refusal is nil, not an error: the buffer already draws a section explaining why a section is empty, and
turning `C-c C-b` into an error would take away the restarts — the decision the buffer exists for — over a missing
annotation.

### The shadow stack, and `backtrace`

plan.org's "Dev vs release builds" table has had **Frames: shadow stack** in the dev column since the project began and
nothing had ever built it. This is it. It is what `(:op "backtrace")` is made of, and it is dev-only, so a shipped game
pays nothing — the same bargain the indirection cells already make.

Chosen over the DWARF route deliberately, and the author's reason is the one that decides it: **the more a break loop
can show, the less often a real debugger is needed.** DWARF buys frames in lldb; this buys them in the break loop,
where someone already is when they want them.

A frame is four words on the calling function's own stack — the frame it displaced, a pointer to a static description
of the function, and two words reserved for locals. The description is per function and not per call, because nothing
about a function changes between two calls to it: the qualified name, `file:line:col` as `Loc` spells it, and how many
slots the frame has. So the name in a backtrace comes off the frame itself and needs no debug information, no symbol
table and no agreement with the optimiser about what a stack frame looks like. It is the same mechanism on wasm32,
which is the other reason it is not `.eh_frame`: plan.org lowers every non-local exit explicitly rather than through
platform unwinding, so there is no unwinder here to borrow.

The compiler emits the push and the pop **inline** rather than calling into the runtime. This is on every call in a dev
build, and a call made to record a call would be most of what it costs.

**The pop is at every `ret`, and the transfer path is the one that matters.** There are five: an explicit `return` with
a value and without, the `none` arm of `(some x)`, the tail of the body, and the landing block a transfer leaves
through. `emit.ml` funnels all five through one `ret`, because the way to get this wrong is to write the pop on the
normal path and not on the other one — and then every *handled* error leaves a dead frame behind, and the backtrace
after the fifth one is five frames of fiction. `test_dev.ml` takes five breaks and resumes all of them by condition
transfer before asking for a backtrace, and the answer has to be two frames. Same lesson `with-allocator` learned about
restoring at the pad.

**A plain global, not a thread-local.** It matches what the handler stack and the restart stack in `flan_rt.c` already
assume: one thread runs Flan, and the listener thread runs C and the loader and never enters a Flan body. If the
language grows threads this becomes thread-local and the compiler's two stores become TLS-relative, which is the whole
of the change.

**The chain is snapshotted on the stopped thread**, into the same `snapshot` the restarts are copied into, at the same
moment and for exactly the same reason: the break loop *polls*, a poll runs Flan, and a chain read by the listener
thread is a chain that can be popped underneath the reader. Names are copied as bytes rather than kept as pointers,
because a transient `C-x C-e` module does get `dlclose`d and its rodata with it. Frame *addresses* are kept beside the
text, because reading a frame's locals means going back to that frame and not to whatever is at index 2 by then.

```
(:op "backtrace")  → (:status "ok" :frames (("fetch" "/game.flan:15:7" "program" 0)
                                            ("main"  "/game.flan:27:7" "program" 0))
                                   :more 0 :stopped t :condition "Missing")
```

Innermost first. `:more` is how many frames deep recursion left off the end — the innermost ones are what the question
is about. **`origin` is `"program"` or `"eval"`**: a break inside a `C-x C-e` thunk has that thunk's frames on top of
the program's, and "where is my program" answered with `eval/7` is true and not the question. The boundary is recorded
at the call, in `flan_agent_poll`, exactly where `restart_floor` is and for the same reason — except that its
out-of-a-thunk value is `-1` rather than `0`, because zero restarts on the stack is a real answer and zero frames
belonging to the program is not.

**Refused while the program is running**, like every other break verb. The chain is the game thread's and it is pushed
and popped on every call; a walk of it from the daemon would have the shape of a backtrace and the contents of a race.

**What it costs, measured rather than assumed.** Two benchmarks, one compiler built per variant, every binary kept and
then run alternately. The figure is the **minimum of nine runs**, because this machine is shared and a mean measures
whatever else was running; the whole table reproduces to three digits on a second pass.

| Dev build, -O2                                            | no frames | frames | frames + slots |
|---|---|---|---|
| 2000 sweeps of a 100×100 grid — sand's inner loop, in miniature | 30.0 ms | 40.0 ms (+33%) | 48.3 ms (+61%) |
| fib(30) plus 20M calls in a loop — nothing but calls       | 28.7 ms | 31.1 ms (+8%) | 35.1 ms (+22%) |

Per call that is about **0.1 ns** for the frame and **0.3 ns** for the frame and the slot table together; per sweep of
the grid, 5 µs and 10 µs, which is 0.03% and 0.06% of a 16.6 ms frame at 60fps. The dev loop's premise is that
redefinition does not stutter a running game, and a sixteenth of a percent of a frame does not. A dev build is already
deliberately slower than a release one — every call goes through a cell, every index is checked, no `defconst` folds —
and this joins that list rather than starting a new one.

The chain's *shape* was chosen by measurement and not before it. An array with a stack pointer — no alloca, no address
escaping — was built and timed as the obvious alternative and is worse on both benchmarks: the frame record on the
calling function's own stack is already hot, and an indexed store into a megabyte of BSS is not. It also has a fixed
depth, which a chain of stack records does not.

### Locals of a stopped frame

The half the shadow stack was built for. `(:op "locals" :frame N)` answers what a stopped frame's named locals hold.

```
(:op "locals" :frame 0) → (:status "ok" :frame "look"
                              :locals (("n" "i64" "3") ("label" "string" "\"hello\"")
                                       ("p" "Point" "(Point {:x 1.5 :y 2.5})")
                                       ("xs" "[3 i32]" "[ 10 20 30]") ("flag" "bool" "true"))
                              :refused (("after" "not bound yet at the point the program stopped")))
```

**Nothing is copied out of the program, and nothing could be.** A Flan value carries no header, so bytes read from
another process would be bytes with no meaning. What the daemon has instead is the *type* — `Tast.fn.slots`, from the
build it owns — and the name beside it in `snames`, which was already there for the DWARF work. So it compiles a thunk
that renders those types **at those addresses, in the program**, on the stopped thread, and reads the text back through
the same seqlocked result buffer `C-x C-e` uses. The only fact that comes from the running program is where the frame
is. That is `render.ml`'s existing walk with its root changed: `Render.render` over `(Deref (Ptr T) (flan/dev-slot f
i))` instead of over an expression — the **pointer-rooted render thunk** NEXT.md said locals needed, and it turned out
to need one new arm in the whole backend (a pointer-to-pointer cast, which under opaque pointers emits nothing).

**A slot's entry is its address, and null until it is bound.** That is the whole of the liveness answer: there is no
analysis, no PC-to-scope map and no bitmap. The store that binds a slot stores the address, so a slot the program has
not reached yet reads as null and is refused by name. Without it, `(let [after 99] …)` sitting past an `error` would
render whatever the stack held, and a slice or a struct of garbage does not misprint — it faults, on the game thread of
a program that is already stopped, which is the worst moment this project has to offer. The daemon asks the program
which slots are bound *before* it builds the thunk, so the set it emits code for is the set the thunk will resolve.

**Only named slots are recorded**, and this is where most of the cost went. A slot whose address is stored anywhere
escapes, and an escaped alloca is one `mem2reg` cannot promote — so recording a slot is paying for it in every call to
that function, for ever. The slots that would hurt most are exactly the ones with nothing to show: `dotimes`'s hidden
bound, the temporaries `(min)` and `(max)` evaluate their operands into, the render walk's own scratch. They keep their
promotion and are refused by name (`s4`, "a slot the compiler made up") rather than shown under an invented one.
Recording every slot instead was built and timed and came out inside the noise on both benchmarks, so the rule stands
on what it shows rather than on what it saves.

**Shadowing is right here, and that is not an accident of this design — it is the thing the DWARF route still owes.**
`check.ml`'s `fresh_slot` only ever allocates, so `(let [v 22] …)` inside `(let [v 11] …)` is two slots, both named
`v`, and both appear with their own values. lldb answers `p v` with 11 in that program and will until a
`!DILexicalBlock` per `Let` exists.

**Four refusals, each by name and with its reason.** Three per slot — invented, not yet bound, and no printer for the
type (a map, a function value, a type variable; the arm exists, no program the checker accepts has reached it yet, so
it is written and untested) — and two whole frames: one belonging to a `C-x C-e` thunk, whose `Tast` the session does
not keep, and one whose *body* is not the body this session holds.

**The superseded frame, and why a count could not find it.** Installing while stopped is deliberately allowed — it is
the fix-it-and-retry loop — so the frame on the stack and the body the session holds can be two bodies of one function.
A slot count catches a body that gained or lost a binding and nothing else; the case that matters is a **rename**, which
changes neither the count nor the types, and which would otherwise show every *new* name against the *old* body's
storage with nothing saying so. That is the confident wrong answer, in the one place someone is working out what went
wrong.

So each `%fninfo` carries a **slot fingerprint**: `Emit.slot_fingerprint` over every slot's name together with the
spelling of its type, since either can change on its own. The frame carries the value for the body it was compiled from,
`flan_dev_frame_slotsig` reads it, the agent snapshots it with the rest of the frame and puts it on the backtrace line,
and `Dev.locals` recomputes it from the body it holds and compares. Different means refused by name — *this frame's body
was redefined since it was entered, so its names no longer describe its values* — rather than answered. The count check
stays in front of it because its message is the more specific one.

Computed in `emit.ml` and read from there by `dev.ml`, so there is one definition of it and the two ends cannot drift.
It stays **off the wire**: `(:op "backtrace")` still answers four fields per frame, because a hash is not something an
editor can act on and the refusal says the fact in words instead. The same is true of the globals fingerprint that
now travels beside it. The bound is worth stating: it is a 30-bit hash, so a
collision is possible in principle, and it would reproduce exactly the silent wrong answer this catches — but only
between two *differing* bodies of a function whose qualified name has already matched, since `find_fn` is what gates the
comparison at all.

The first build of this shipped the fingerprint into `%fninfo` and stopped there — no accessor, no wire field, no
comparison — and was left in the tree with its own test red. Four of five hand-offs missing looks exactly like one
hand-off dropping a number, which is what the note left behind said it was.

**A `(Vec T)` shows as `<vec>` and a `(Ptr T)` as `<ptr>`**, because that is what `render.ml` already does for them
everywhere else: following a pointer a REPL was handed is not a safe thing to do on someone's behalf, and walking a
`Vec` structurally is a walk over storage the frame does not own — `(print (as-slice v))` is how that is asked for,
and it says at the call site that it borrowed.

Each slot is rendered **from its address** rather than copied into the thunk first. A copy would be one `alloca` the
size of the slot — 40KB for sand's grid — and the walk only ever shows eight elements of it. The cost is one call to
`flan/dev-slot` per leaf the walk reaches rather than one per slot, which the depth and span caps already bound.

### Globals of a stopped stack

The other half, and in this language arguably the more useful one: a game keeps most of its state in top-level
`defvar`s and `sand.flan` holds its entire grid that way, so "what is this program's state right now" was a question
with nowhere to ask it.

```
(:op "globals") → (:status "ok"
                      :globals (("grid" "[4 i32]" "[ 7 5 0 0]" (0 1))
                                ("pressure" "i64" "12" (0))
                                ("label" "string" "\"running\"" (1)))
                      :refused () :skipped ())
```

**Not per frame, and that is the design rather than a layout preference.** A global is not part of a frame — it is
program state the frame happened to touch — so nesting it under one implies an ownership that is not there and repeats
the name once per frame that reads it. So: one section, whose contents are the **union of the globals every frame on
the current stack references**.

**The compiler does the choosing.** `Reach.expr_refs` is the walk that already computes what a function refers to — it
is how the link drops a package nothing calls — and pointed at one body it answers that body's reference set. Listing
*all* of a program's globals instead would bury the one that matters under the prelude's PRNG state; taking only what
the stack reaches is a filter the compiler can apply and a person cannot. Direct references only, with no transitive
closure through calls: a callee that reads a global is either on this stack, contributing its own references already,
or it is not on it and is not part of where the program stopped.

**Each entry carries the frames that touch it, by index**, which is what the conditions buffer already numbers frames
by. That recovers what per-frame nesting would have told you — "the whole chain is reading this" reads differently from
"only the innermost does" — at no cost in duplication. **Ordered by the innermost frame that touches it**, because a
deep stack makes the union large and proximity to the error is what puts the likely culprit on top; ties keep
declaration order.

**The mechanism is one step simpler than `locals`.** A local is reached by *address*, and only the stopped program
knows where its frame is — hence `flan/dev-slot`, hence the `bound_slots` round trip, hence the null-until-bound
refusal. A global is reached by *name*: `Emit.redefinition` writes a global the host already has as `external`, so the
dlopened thunk binds to the program's own storage and the dynamic linker does the work. Nothing is asked of the stopped
thread at all, and there is no not-yet-bound case, because a global's storage exists from the moment the process
started. It is `Render.render` over `(Global n)` — the same root `C-x C-e` uses when you type a global's name at it.

**A frame that cannot be attributed contributes nothing and says so.** An `eval` frame has no declaration in this
session; a lifted handler clause has none of its own; a frame whose body was redefined since it was entered holds a
body whose reference set is a claim about different code. All three go into `:skipped` with the reason, because "the
union is incomplete and here is why" and "these are all of them" are different answers and the second one is the lie.

**The hole this had, and the second fingerprint that closes it.** `Emit.slot_fingerprint` hashes a body's *slots*. For
`locals` that is exactly the right cut: identical slots means the names still describe the storage, so the answer is
still true. Here it was not, because a body can change which globals it names without touching a single slot — and then
this section showed the *new* body's reference set attributed to the *old* frame. The values were never at risk; they
are read from the program's storage by name. What was, is one frame's membership in the union and the frame numbers
beside an entry.

So `%fninfo` carries a second number beside the slot fingerprint: `Reach.ref_fingerprint`, over the **set** of globals
the body names — sorted and deduplicated, because a reference set is not ordered and a body that mentions the same two
globals the other way round is the same body, where slot indices make the slot fingerprint order-sensitive on purpose.
It travels the same path the first one does: `flan_dev_frame_refsig`, the agent's snapshot, the backtrace line, and
`Dev.globals_op` recomputing it from the body it holds. Different means that frame is in `:skipped` with its own
reason — *this frame's body names different globals than the one this session holds* — and the rest of the stack still
contributes.

**Two numbers and not one combined**, which is the whole reason this is a second fingerprint rather than a wider first
one. They are different facts: a frame whose slots match and whose globals do not has locals that are perfectly
readable and attribution that is not, and one hash over both would make `locals` refuse a frame with nothing wrong
with it. So `locals` still checks the slot fingerprint alone.

It lives in `reach.ml` rather than `emit.ml` because `Reach.expr_refs` is already the walk that answers "what does this
body refer to" and is the same walk the union above is built from — the two ends cannot disagree about what counts as a
reference. Which names are globals is the caller's to say: the emitter knows the program's globals, and so does the
session. `test_dev.ml` now drives the exact case — a redefinition that binds identical locals and names `untouched`
where the stopped frame's body names `pressure` — and with the check disabled it fails twice: once on the missing
refusal, once on `untouched` appearing in the union under frame 0.

### Conditions — step 2: `restart-case` and `invoke-restart`

`spec-conditions.md` §3 to §6: the transfer. A handler runs where the signal was, decides, and control resumes at a
`restart-case` further out.

```
(defn fetch [n i32] i32
  (restart-case (middle n)          ; its value if nothing transfers
    (use-placeholder [] -1)
    (retry           [] 7)))

(handler-bind [(AssetMissing [c] (invoke-restart 'use-placeholder))]
  (fetch 2))                        ; -1
```

**The channel is an out-parameter**, as §6 now says: one `ptr` appended to every Flan signature, written by an
`invoke-restart` and checked after every call. The return type stays what the source says, so the disassembly is the
release one plus a guard, and one pointer threads down the whole chain — a callee writes the target into its caller's
slot and each frame only has to check and return early, which reuses the existing `return` path and with it §5's defers.
`Emit.signature` was already the one place a signature is spelled, which is what made this a three-line change rather
than a hunt.

**Every function is transfer-transparent**, release included — and that is the ABI, not a stopgap. A cell holds a bare
pointer, so the honest answer to "what can this call?" is "anything"; the same bargain as the indirect call. §6 and
plan.org both now say the later optimisation may stop a function *checking* the channel, or pass the pointer straight
through, but may not drop the parameter — a signature that depended on an analysis could not be reloaded into. Uniform
also means redefinition acquires no new refusal class.

**The transfer target is the restart frame's own address, not a clause id.** This is a correction to what the previous
note settled. A static id has to be unique against every module a running program may *later* load, and a hash is only
probably unique — two `restart-case`s colliding means the inner one silently catches a transfer aimed at the outer. The
frame is an `alloca` in the function that offers it, so its address is exact, and it also says *which* clause, which is
how clause ids disappeared entirely. §6 says "transferring to frame N" and this is closer to it than the number was.
Re-entering a `restart-case` then works with nothing extra: each activation allocates its own frames, and §4's
"innermost offering the name" is just the order of the walk.

**Cleanup happens in landing blocks, one per region.** A guard branches to the innermost open one, which pops whatever
frames it established and either catches the transfer or forwards it outward:

- a `restart-case`'s pops its restart frames, compares the target against its own, and either runs that clause or puts
the target back and goes on out;
- a `handler-bind`'s pops its handler frames and goes on out — which is the path a transfer out of a handled body takes,
and without it the handler stack would be left pointing into a frame that has gone;
- the function's own runs its defers (§5) and returns early. `errdefer` does not run and never could: `try`/`Result` is
still refused by name.

A single function-wide unwind block would have been wrong for the first two: a call inside a `restart-case` body would
jump straight past the very form that was supposed to catch it.

**The channel is cleared before any cleanup runs and put back after.** A defer makes ordinary calls and each one is
guarded; with the channel still set the first of them would branch straight back into the landing block it came from.
Same reason the clause body starts with it null.

`flan_signal` takes the channel and passes it to each handler, and stops walking once one has written to it. That makes
the one C frame every handler is reached through transparent to a transfer — it has to be, or §6's "a transfer cannot
cross a foreign frame" would make `restart-case` useless. It is also the only such frame: `extern` is Flan-to-C only and
there are no function values yet, so nothing can call *back* into Flan across one.

Scope, each piece refused by name with its reason and a test on the reason:

- **restarts took no parameters.** That covered §1's own `load-texture` example and skipped argument marshalling and
§3's runtime check. Both are in now; see "Conditions — step 3" below.
- `return` inside a `restart-case` body, exactly as inside `handler-bind`: a bare `ret` skips the pops.
- one `restart-case` offering a name twice — §4 finds the first frame offering it, and two in one frame makes that a
choice nothing in the source shows.
- `invoke-restart` inside a `defer`. A defer *is* the cleanup a transfer runs on its way out, so a transfer starting
there leaves the function's defers half run with two targets and no way to choose. The lexical case is the checker's; a
defer that reaches one through a call is trapped at run time by `flan_transfer_fail`, because nothing static could see
it.
- no restart of that name is active: a runtime error at the invoke site, named and located, rather than an unwind past
everything. There is nowhere to resume, so there is nothing else to do.

**`(error c)`, §2.** The same walk as `signal`, and the difference is entirely what happens when the walk ends: `signal`
returns `Unit` and the signalling function carries on, `error` has type `Never` and stops. So only a transfer gets past
it, which is why `emit` puts a guard after the call and then `unreachable` — and why `flan_error` cannot be marked
`noreturn`, since it does return, on exactly one path. Being `Never` is also what lets it stand as a `restart-case`
body's fall-through, which is the shape §1's `load-texture` example needs. `test/programs/error.flan` is the unhandled
case: it cannot be an `outputs` row, because it does not exit 0.

`flan_transfer_fail` covers the ordinary return path as well as the unwind one: a defer that reaches an `invoke-restart`
through a call traps either way, and the message names the rule rather than the path, since the rule is the same.

**A lifted handler clause is named after the function it came out of** — `handler/step/0/Missing` — and is emitted by a
redefinition module alongside the body it belongs to, hidden, for the same interposition reason the body is. This was a
hole step 1 left: the name used to be numbered by position in the whole program's lifted list, so it was neither stable
nor attributable, and a redefinition of a function containing a `handler-bind` failed in `llc` with an undefined value.
A clause is reached by address from its parent's body and from nowhere else, so it takes no cell and no registry slot.

`test/test_dev.ml` drives that path: a third evaluation redefines `step` to a `restart-case` whose frame is an `alloca`
in the newly loaded module, whose guarded call goes through the host's cell, and whose transfer starts in a handler and
crosses `probe`, which the host was compiled with. Those three do not meet anywhere else.

Two things found by writing it:

- **`{ ctx with in_handler = true }` was a latent bug.** `ctx.slots` and `ctx.slot_tys` are mutable, so a copy allocates
the body's slots into a record the function never sees again and the indices collide. It was harmless only because no
`handler-bind` body in the tests had a `let` in it. The flags are set on `ctx` and restored now.
- **`test/reload_host.c` had to learn the parameter.** It calls `flan.outer` through an `__asm__` label, which does not
fail at link time when the prototype is a parameter short — it reads a garbage pointer as the channel and dies somewhere
else entirely.

`test/programs/restarts.flan` runs in the acceptance table at `-O2`, at `-O0` and as a dev build. `-O0` is not redundant
here: the guard after every call is control flow the optimiser would otherwise launder, and the dev build is where each
of those calls goes through a cell.

### Conditions — step 3: restarts take parameters

§3's other half, and the half the comparative studies all ask for: `use-value` and `store-value` are the two restarts
whose answer is not in the program.

```
(restart-case (middle n)
  (use-value [v i32] (* v 2))
  (retry     []      7))

(handler-bind [(AssetMissing [c] (invoke-restart 'use-value 21))] (supplied 7))   ; 42
```

**The parameters live in a buffer the restart-case owns.** The obvious place is the invoker's frame — it is where the
values are — and it is wrong: a clause runs after every frame between the invoke and the target has returned (§5), so
the invoking side is gone by then. The invoker stores into the *target's* buffer while both are still alive, which is
the one moment they are. A clause's parameters are then ordinary slots of the establishing function, loaded out of that
buffer in the clause's landing block, and the clause body is in-frame code that sees this function's scope like any
other.

**What a restart takes is checked at run time, and it has to be.** §4 finds a restart by name on a dynamic stack: the
invoke site cannot see what it will find and the frame cannot see who will find it, so there is nothing for the checker
to compare. The frame therefore carries its parameter count and a 32-bit hash of how the types are spelled, and
`invoke-restart` compares both against its own before it stores anything. Every frame carries them, parameterless ones
included — a clause taking none has to refuse arguments as loudly as one taking two of the wrong type. The count is not
redundant with the hash: it is what keeps a hash collision between two *different* signatures harmless, since a
collision would then have to be between two lists of the same length. The spelling itself rides in the frame as well,
because the message has to say what was wanted and what was given and neither end knows both:

```
restarts.flan:79:36: restart use-value takes (i32), given ()
restarts.flan:87:36: restart retry takes (), given (i32)
```

**The arguments are evaluated into slots before the invoke node, not hung off it.** Two reasons and both are real. An
argument can transfer on its own, and that guard must fire before anything aims the channel. And a call written inside
an argument has to be on the walk `Reach` and `Load` already do — `InvokeRestart` was a leaf to both, and a leaf that
grew a subexpression would have dropped a function that is called from nowhere else and failed in the linker.
`restarts.flan` has exactly that function, `half`, to keep the claim tested.

**The break loop can take a restart it cannot fill in, so it is refused.** A transfer has two sources: an
`invoke-restart`, which writes the arguments first, and the break loop, which aims the channel at a frame by position
and has no value to supply. They reach a clause through the same channel by design, so nothing downstream can tell them
apart — which is what makes this the kind of hole that ships silently. The frame is pushed with its buffer marked
unfilled, `invoke-restart` marks it filled, and a clause with parameters checks the mark before reading. Choosing
`use-value` from a break loop today stops the program and says why. Filling it in is the editor half, and it is now the
top of `NEXT.md`: the answer is a Flan expression, and there is already something that compiles one against the live
program.

**Lookup stayed by name, and the signature is checked against what it found.** `flan_find_restart` matches the name
hash and nothing else, so an inner `(use-value [s string] ...)` shadows an outer `(use-value [v i32] ...)` and an i32
is refused there — the outer frame that would have taken it is never consulted. That is §4 read straight ("the first
frame offering the name") and it is the thing a reader will assume works the other way, so `restarts.flan` has a case
for it. Searching outward for a frame whose signature fits would make which restart runs depend on the arguments,
which is overload resolution on a dynamic stack.

Both of these guards go through `fail_block` unconditionally, unlike the bounds checks: `--no-bounds-checks` does not
remove them. A wrong index is a wrong answer, and a transfer into a clause whose parameters were never written, or
written to a different layout, is memory corruption.

`runtime/flan_rt.c` gained two message functions and nothing else. The restart frame's first four fields are the ones C
declares and their offsets do not move; everything §3 needed is appended after them, and C never allocates one.

§3's other open point, a **report string per clause**, is still open and was not settled first as §3 asks. The field
and the accessor are both cheap; the only thing that would read them is the break loop's listing, which lives in the
agent and the daemon, so it would have landed as a field nothing read. It goes with the editor half, which is changing
that listing anyway.

### What is left
- **Editor comforts**: completion, eldoc, jump-to-definition, error overlays.

**Session identity is the daemon that owns the build.** A session's struct layouts and global types have to describe the
memory of the process it is talking to, which is only guaranteed if it is the session that compiled the running binary.
Attaching to a process someone else built is not a thing to support by default.

## Allocators, `(Vec T)` and `StorageExhausted`

`spec-memory.md`'s Allocators section is frozen and settles what to build. This is why the built thing is the shape it
is, and — separately and loudly — the three places it **amends** that section plus the one thing it adds to it.

### `Allocator` is a builtin opaque type, so none of milestone 5 was needed

The spec defines an allocator as "a procedure plus an opaque data pointer", which reads as a function value, and
`check.ml` refuses function values four ways — a written `(Fn ...)` annotation, a written `fn` literal, a `defn`'s name
in value position, and calling anything other than a named function. Read straight off those lines, containers need
function values and the work doubles.

None of the four is anywhere near this. All four are about *surface syntax*, and a value the compiler builds that no
surface form names trips none of them:

- `Allocator` is a `Types.t` case with no user-writable constructor, the way `string` is a builtin ptr+len. No Flan
  type names its procedure.
- The procedure is a C symbol the emitter names.
- `vec-new`, `push`, `free`, `free-all` and the rest are ordinary named calls, which `check_call` already routes
  through `named_call`.

The precedent was already in the repo twice: a `handler-bind` clause is lowered to a function called back through
`h->fn(condition, xfer)` and built as its own `Tast.fn`, never an `Ast.Fn`; and a dev build's `call` loads a pointer
out of a cell and calls through it, which is the indirect call the surface language refuses.

What *does* need milestone 5 is a **user-written** allocator: "here is my proc, make an `Allocator` from it" wants a
`defn`'s name in value position. `make-allocator`, `allocator-from` and `allocator` are refused by name with that
reason, rather than coming back as unknown functions.

**An `Allocator` value is a pointer to the runtime's struct, never a copy of one.** That is forced, not chosen. The
capability set has to be readable from wherever a container landed, and `free-all` bumps an epoch every container made
from the allocator has to observe. A copied-by-value allocator gives each copy its own epoch and the dev trap never
fires.

### The surface

| Name | What |
|---|---|
| `(heap-allocator)` | the general-purpose tier: malloc, aligned, with free |
| `(arena-new bytes)` | a bump arena over one fixed backing buffer |
| `(arena-destroy a)` | hands the pages back — see the amendment below |
| `(free-all a)` | releases everything the allocator holds, retain-capacity |
| `(can-free? a)` / `(can-free-all? a)` | the capability set, read at run time |
| `(alloc-epoch a)` / `(alloc-id a)` | the counter `free-all` bumps; the allocator's identity |
| `(alloc-budget a)` / `(set-alloc-budget a n)` | a ceiling on live bytes — an addition, see below |
| `(alloc-live-blocks a)` | "did you forget to free", answered at the tier that can answer it |
| `context/allocator` / `context/temp` | the current implicit allocator, and the per-frame arena |
| `(with-allocator a body...)` | rebinds for a dynamic extent and releases nothing |
| `(vec-new T)` / `(vec-new T a)` / `(vec-new)` | a Vec, against the context or a named allocator |
| `(push v x)` / `(reserve v n)` | `Unit`, both |
| `(at v i)` / `(len v)` | the array names, extended — not a parallel pair |
| `(as-slice v)` / `(as-slice v lo hi)` | a non-owning `[T]` view |
| `(clone v)` / `(clone v a)` | the only copy; assignment moves |
| `(free v)` | consumes its argument |

### Three amendments to a frozen spec, and one addition

**1. `free-all` is retain-capacity, and `arena-destroy` is the operation that hands pages back.** The spec's table has
`free-all` and nothing else. Zig's `ArenaAllocator.reset` takes a `ResetMode` of `free_all` / `retain_capacity` /
`retain_with_limit`; Odin's `arena_free_all` is retain-capacity in effect, because its arena is one fixed backing
buffer and the call only sets `offset = 0`. For a frame arena reset every frame, retain-capacity is the normal case and
handing the pages back only to ask for them again is the unusual one. Taking the mode as a parameter would have grown
the table the spec froze at four names; a second operation does not. The epoch is bumped either way — the pages being
the same does not make a container built before the reset valid, which is the whole point of the trap.

**2. `context/allocator` and `context/temp` are dynamic variables with save and restore, not extra parameters.** The
spec says the allocator is "part of the calling convention". The literal reading touches every function signature, the
FFI shim, the dev trampolines and the reload ABI, for the same observable behaviour, and it collides with every other
lane working in `emit.ml`. The dynamic variable is the implementation; the literal reading is deferred and is a
performance question (a parameter avoids a global load), not a semantic one.

**3. The `Vec` header is six words in every build, not four in release.** The spec fixes the release layout at
`ptr + len + cap + allocator` with the generation word and the epoch dev-only. A layout that changes with a build flag
is a layout that can disagree *silently* across the reload boundary: a redefinition module is built by `llc` and `ld`
against a host built separately, and nothing makes the two agree on a struct size. So the words are unconditional and
so is the epoch check. The 32-byte release layout is deferred on that, and it needs the reload path to carry the flag
before it can land.

**The addition: a budget.** `flan_allocator` grew a ceiling on live bytes, 0 for none. The spec's `retry` restart is
answerable only by a handler that can make the *same* request succeed, and for a fixed backing store the handler that
works is the one that raises the ceiling — releasing the region the container lives in invalidates the container, which
is exactly what the epoch check catches. The spec names "grows the arena and then invokes `retry`" as the handler that
works; something has to be growable for that sentence to be true. It doubles as how a test exhausts an allocator on
purpose.

### `with-allocator` is its own IR node because of the transfer path

Save, run, restore — and *restore again at the pad*. That second restore is the whole reason it is a node rather than a
`let` and two calls. A body that errors, or one a handler transfers out of, leaves through `current_pad`, and a context
allocator left pointing into a region nobody outside the body has heard of would be wrong in the break loop, which is
precisely where something is about to allocate in order to render a condition. `test/programs/allocators.flan` asserts
that path by taking a restart out of a `with-allocator` body.

It releases nothing, per the spec: not at the end of a `let`, not at the end of a function, not at the end of the body.
The program proves it by reading the epoch either side.

### `(Vec T)`: one type-erased runtime, and the backend learned almost nothing

The element type appears nowhere below the call site. `size_of` and `align_of` are produced where the concrete type is
known — which without generics is simply the concrete call site — and passed in, which is Odin's arrangement
(`base/runtime/dynamic_array_internal.odin`). The backend grew four prims in total:

- `Rt of string` — a call into the runtime's C named by symbol, with argument and result LLVM types read off the
  expression nodes. The container runtime is type-erased and therefore *is* a list of C entry points, so one arm covers
  all of them. A `Vec` argument crosses as its address, which is also what lets an operation mutate the caller's Vec.
- `SizeOf` / `AlignOf` of a type, filled in from the same layout calculator DWARF uses — the one the acceptance test
  already checks against LLVM's own `getelementptr` answers.
- `AddrOf` of any expression, place or not, because the element a `push` copies may be computed. The backend already
  spilled a non-place to a temporary for exactly this.

`at` and `len` were already the names for a fixed array and a slice, so a Vec extends them rather than adding a
parallel pair — the asymmetry `nth` was removed for. The value form `(at v i)` and the place form `(set (at v i) x)` go
through one helper, so they cannot drift apart the way `nth` did.

A Vec's length and index are `i32`, like every other length here. Widening indices is one change across every
container and not a Vec question.

`let` has no type annotation — `parse.ml` settles that a triple binding is ambiguous and that types are inferred — so a
local Vec has nowhere to say what it holds and the element type is written at the call: `(vec-new i32)`. This is *not*
the explicit instantiation syntax the generics section rules out: nothing here is generic, and the name resolves as an
ordinary type rather than binding a type variable. Where the context does say — a `defvar`'s type, a return type, an
argument — it may be left out.

A zeroed Vec has a null allocator, and the first operation that needs storage adopts the context allocator, which is
Odin's behaviour. The alternative was refusing a Vec-typed struct field outright; that is refused anyway, for a
different reason (below), but the adopt rule is what makes `Zero` of a Vec a usable value rather than a null deref.

### Move-only is a dead set, and it is flow-sensitive at a join

Reading a move-only local is a move unless the site said it was a borrow. That is the conservative direction: passing
one to a function, binding it, returning it and `free`ing it are all moves and all reach one place, and the handful of
operations that only look at a container (`at`, `len`, `as-slice`, `push`, `reserve`, `clone`) say so. Only a
*syntactically simple* target counts as a borrow — in `(len (f v))` the call still moves `v`.

At an `if` and at a `match`, every arm is checked from the state before the form and the **union** of what they moved
survives the join. A flat set is wrong in both directions: it refuses `(if c (free v) (free v))`, which is legal, and it
accepts a use after a one-armed move, which is a use-after-free. The arms are alternatives, and that is what a union
says.

The one case a dead set cannot answer is a move inside a loop: merged once at the end of the body it counts one move,
not two, while the second iteration would use what the first gave away. So it is a rule rather than an inference — a
move of a binding declared outside the loop is refused, with that as the reason.

### What ownership is not transitive through yet, and why each is refused

The spec says ownership is structural — a struct containing a Vec is itself move-only, `free` recurses into owning
fields, and a field cannot be freed on its own. That machinery is the recursive teardown `drop` brings. Until it lands,
three shapes are refused where they are declared, each naming `drop`:

- **a struct field of `Vec` type**, because the struct copies its header on assignment and nothing records a move;
- **a global of `Vec` type**, because the dead set is per function — two functions each freeing it is a double free
  nothing could see, and a global read does not go through the move path at all, so even the one-function case would be
  accepted. Half a rule is worse than none. A global **`Allocator`** is a different thing and stays legal: an allocator
  is a copyable opaque handle, and it is what makes a handler that owns the arena expressible, since a handler cannot
  see the locals of the function that established it;
- **a `Vec` of a `Vec`**, because the type-erased runtime copies and releases elements bytewise: `clone` would
  duplicate inner headers instead of copying what they own and `free` would drop their buffers.

And a `Vec` does not cross to C: handing a header that owns storage to C hands out an owner. `(as-slice v)` as
`(Ptr T)` plus `(len v)` is the shape that does cross, and the refusal says so.

### `StorageExhausted` went in *with* `Vec`, not after it

No allocating operation returns an error and none can fail silently. When the allocator cannot satisfy a request the
operation signals `(StorageExhausted {.bytes n .align a .allocator id})` with `error` — whose type is `Never` — inside a
`restart-case` offering `retry`. One rule over every allocating operation, which is what keeps `push` and `reserve` at
`Unit`, `clone` at the container, and no signature anywhere growing a `Result`.

It had to land with step 2 rather than after it: retrofitting adds a transfer check to every call site of every
allocating operation, which is the point of having decided it first. Odin's `append` returns an ignorable
`Allocator_Error`, and its type-erased path returns the old length on a failed reserve; an append that appends nothing
and says nothing is the outcome this rule exists to make impossible.

The lowering is built out of nodes that already existed, so the backend learned nothing about allocation:

```
(let [ok false]
  (while (not ok)
    (restart-case
      (do (set ok ATTEMPT)
          (if (not ok) (error (StorageExhausted {...}))))
      (retry []))))
```

A handler that frees something, releases a scratch region or raises the ceiling and then invokes `retry` lands in the
clause, the clause falls through, and the `while` re-attempts the **same** request. Every argument to the attempt is
bound to a slot before the loop, so a retry re-attempts the allocation and not the expression that produced the value a
`push` was given. With nothing handling it, `error` stops the program on the frame that erred.

This is the named exception to plan.org's "restarts go at the resync point, once" — the restart is established at the
failing allocation, because a restart at some outer loop cannot re-attempt an allocation and only the allocation site
can. The condition is a value struct with fixed numeric fields and **no rendered message**, because formatting would
allocate and this is the one path that must not; the numbers of the failed request are read back out of the runtime.

### The epoch trap, which is the shipping answer to an open question

`spec-memory.md` leaves "catching a use-after-release statically" open on purpose: the static rule needs to know which
allocator a construction used, and `with-allocator` plus `context/allocator` are exactly the mechanisms that deny that
knowledge. The shipping answer is detection. A Vec records the epoch of the allocator it was made with, `free-all` bumps
that counter, and any operation on a container whose recorded epoch has moved traps naming the site.
`test/programs/stale-region.flan` is the case, and the point of it is that `v` is still in scope, still looks fine, and
nothing marked it — which is precisely what a static rule cannot see.

**The generation word has no reader.** It is bumped on every reallocation, as specified, and the stale-slice trap it
exists for is not implemented: a slice is ptr+len and has nowhere to carry the Vec's identity or its generation. Said
plainly here rather than implied by the word's presence in the header.

### What this leaves for steps 5 to 7

`(Map K V)` is built — see below. What is left: `drop` and with it the transitive move-only rule, recursive teardown,
and the refusal to construct a drop-carrying container against an allocator without `can-free`; `(Result T E)` and
`try`; generics; the macro expander.
And the **accumulation pattern** — `(fn [c] (push errors c) ...)` over an enclosing Vec — which `Vec` does not buy:
capture does not exist at all, and the spec's captured-`Vec`-by-pointer rule has never had to exist because every
capturable type today is a value type. It is its own item and should be planned as one.

## `(Map K V)`, which is Odin's map

Step 4 of the container build order, over the runtime `Vec` already established. `spec-memory.md` is followed as
written; the two places this departs from *Odin* are named below, and there is one amendment to the spec and one
restriction under it.

### The three properties that were the point

Odin's header states them and they are why it was the thing to follow (`base/runtime/dynamic_map_internal.odin`).

- **Open-addressed Robin Hood hashing at a 75% load factor.** No buckets and no per-entry allocation: one block holds
  every key, every value and every hash. Robin Hood is the part that earns its keep — on insert, an element that is
  further from the slot its hash wanted than the occupant it is looking at takes the slot and sends the occupant on.
  Probe distances even out, and a lookup may stop the moment it is further from home than the occupant it is looking
  at, because no element is ever further from home than one it passed. That early exit is why a miss costs about what
  a hit does.
- **Cache-line cell packing.** A flat `[capacity]K` array lets one key straddle two cache lines, so a probe walking
  four slots can touch five lines. A cell packs as many `K` as fit in 64 bytes and pads the rest, so no key ever
  straddles a line. Keys, values and hashes are three separate runs, so a probe — which reads hashes, and only then
  one key — touches hash lines and nothing else until it has a candidate.
- **Pointer-width integers throughout**, so no sign extension or masking gets into the probe loop.

### Two departures from Odin, both deliberate

**There are no tombstones**, because `spec-memory.md` defers removal ("Move-aware lookup, removal, and owned entries
are deferred"). A slot is empty or occupied and nothing else, which deletes Odin's backward-shift loop entirely — it
is the single largest reason this file is shorter than the original. When removal arrives, that loop is what it costs.

**The header does not tag the capacity into the data pointer.** Odin stuffs `log2cap` into the low six bits because
its `Raw_Map` must be three words. This header already carries an allocator, a generation and an epoch, so the tagging
would buy nothing, cost a mask on every access, and — the part that actually matters — make correctness depend on the
block being 64-byte aligned. Alignment is *requested*; an arena whose base is not cache-aligned now gives a slower map
rather than a wrong one.

The header is six words, 48 bytes, the same as a `Vec`'s and for the same reason: a layout that changes with a build
flag can disagree across the reload boundary.

    data  len  log2cap  allocator  gen  epoch

### The hash and equality pair, and why most key types do not get one

`spec-memory.md` restricts keys to built-in structural types and makes hashing and equality compiler-provided
structural operations. So there is no dispatch to design: a key type resolves to a pair of symbols, passed to the
type-erased runtime the way Odin hangs two contextless `proc`s off a `Map_Info`.

Most key types need nothing emitted. A key whose equality is bytewise and whose bytes are all present — every integer,
enum, bool, and fixed array of those — is served by one runtime pair over `(pointer, size)`. Two kinds are not, and
the reasons are worth keeping because they are what a bytewise shortcut would have got wrong:

- a **string** is ptr+len and its bytes are elsewhere, so two equal strings at different addresses must hash alike;
- a **struct** may have padding, whose bytes are indeterminate — two structs equal field by field can differ bytewise
  — and it may hold a string, which brings the first problem inside it.

A struct therefore gets a pair emitted for it, walking its fields in declaration order and addressing nothing but
fields. Two maps with the same key type share one pair, and a struct reached twice through two fields emits one.

`Tast.FnAddr` is what carries the symbol, and it is **not** a function value: nothing in the surface language can
produce one, name its type, or call through it. It is the same escape the allocator used — the one NEXT.md predicted
would work — and `reach.ml` learned the edge, because a function reached only by address is invisible to the
reachability walk otherwise. That was already true of handler clauses.

A pair emitted for a struct is an ordinary Flan function, so its signature ends with the transfer channel like every
other. The runtime's typedef spells that pointer out rather than hoping nothing writes through it, and each built-in
hasher exists in two spellings — the pointer form that matches the typedef, and the direct form an emitted hasher
calls per field, which has no channel to hand on.

### The surface

| Name | What |
|---|---|
| `(map-new)` `(map-new K V)` `(map-new a)` `(map-new K V a)` | a new map; the pair may be omitted where the context says |
| `(put m k v)` | upsert, `Unit` |
| `(get m k)` | `(Option V)` — absence is `None` |
| `(has-key? m k)` | `bool`, copying no value — **an addition; the spec does not name it** |
| `(len m)` `(reserve m n)` `(clone m)` `(clone m a)` `(free m)` | extended, not duplicated |

`has-key?` is **not in `spec-memory.md`** and is an addition, flagged because everything else here is the spec's.
`(get m k)` answers the same question, but through an `Option` the caller then has to match, and the common use is a
condition. It copies no value, which is also why it is not just `get` with the result thrown away. The `?` suffix
follows `can-free?`.

`len`, `reserve`, `clone` and `free` were **extended rather than given map-shaped names of their own**, which is what
`at` and `len` already did for `Vec`: one question, one word. `reserve`'s `n` is entries, not slots — the runtime sizes
the block so `n` still sits under the load factor, which is the only reading of "room for n" that does not reallocate
on the nth put.

`get` builds the `Option` in the checker, not the runtime, which has no idea what an `Option`'s layout is; keeping it
that way is what lets one entry point serve every value type. `put` and `get` bind their arguments to slots before the
allocation guard, so a `retry` re-attempts the allocation and not the expressions that produced the key and value.

**`{K V}` is the type spelling and there is no map literal.** A bare map form in expression position is a struct
literal's field list, and giving the same braces two meanings is what the colon-to-dot change was for. A map is built
with `map-new` and filled with `put`.

### The refusals, each by name

- A **float key** — not a milestone question, which is why it is said separately. NaN is not equal to itself, and
  `0.0` and `-0.0` are equal while differing bytewise. There is no equality there for a map to hash.
- A **`Ptr`, slice, `Vec` or `Map` key** — would hash an address rather than what it points at, which is a different
  operation.
- A **fixed array whose elements are not compared bytewise** — an array of structs or of strings needs the
  per-element walk a struct key gets, driven by a loop rather than a field list. Nothing has wanted one, so it is
  refused rather than written untested, and the shape that does work (a struct holding the array) is named beside it.
  **This is narrower than the spec**, which lists fixed arrays without qualification.
- A **move-only value** — the refusal `(Vec (Vec T))` already carries, for the identical reason: the runtime copies
  entries bytewise, so `clone` would duplicate headers and `free` would leak what they own. Owned entries arrive with
  `drop`.
- **`Unit` as a value** — there is nothing to store, and the cell geometry divides a cache line by the element size.
  Named rather than left to divide by zero, because it is the natural spelling of a set.

`StorageExhausted` under `retry` holds over `map-new`, `put`, `reserve` and `clone`, reusing the machinery that landed
with `Vec`; no operation returns an error. A map is the harder of the two containers for that rule, and
`map-exhausted.flan` is why it gets its own program: a `Vec`'s failing allocation leaves the `Vec` untouched, whereas a
map's growth allocates a new block, rehashes into it, and only then releases the old one — so a failure partway must
leave the map exactly as it was, or the retry re-attempts against a half-moved map.

The epoch trap covers the map too, and `map-stale-region.flan` is separate from `stale-region.flan` because the two
reach the check by different routes: a `Vec`'s operations check on the way in and stop there, while a map's `get` goes
on to call a hash and an equality function through pointers into the block. A missing check there is not a wrong
number, it is a probe loop walking released memory.

### Is this another Python dict? — measured

The question was asked directly and the answer is that **Python's algorithm is fine**. What makes CPython's dict slow
is that every key and every value is a separately allocated, reference-counted object and hashing goes through
`__hash__` and `__eq__` calls that cannot be inlined. This map stores raw bytes in the block and compiles hashing and
comparison concretely per key type. That is most of the gap before any cleverness.

Measured on this machine, `i64` to `i64`, against CPython 3.13's dict on the same workload:

| Working set | Flan | CPython dict |
|---|---|---|
| 10k entries, 10M lookups (cache-resident) | 21 ns/lookup | 132 ns/lookup |
| 1M entries, 10M puts + 10M lookups | 1.41 s | 1.16 s |

**Six times quicker cache-resident, and slower at a million entries** — and the second row is written down rather than
left out. Both are waiting on memory there, and this layout waits longer: keys, values and hashes are three separate
runs, so a lookup that misses everything takes three cache misses where a compact dict takes two, and the hash run is a
full eight bytes a slot. Cell packing buys probe locality, which is a win while the hash run is resident and a loss
once nothing is. One byte of metadata a slot — the Swiss-table arrangement — is the known answer and is not built.

The path from 35 ns to 18 ns (the cache-resident floor, at 500 entries) was **profiled, and the first two guesses were
both wrong**: the per-slot cell division and the block-size divisions were each replaced first and neither moved the
number. What did: FNV one byte at a time became eight, and a key that is one machine word became one load and one mix
with no loop; equality on eight bytes stopped being a call into libc's vectorised `memcmp`, and copying a value out
stopped being a call into `memmove`; the block geometry stopped being recomputed five times over in a function that ran
twice per lookup; and the seed stopped being a five-multiply avalanche on the critical path for mixing the hasher does
again immediately after. `64/size` is a table, which is Odin's `Map_Cell_Info` by another route — Odin precomputes it
per type because the probe loop must not divide, and here the sizes arrive as ordinary arguments.

What remains at 18 ns is the type erasure itself: a non-inlinable call into the runtime and two non-inlinable indirect
calls to the pair. That is the trade `spec-memory.md` chose deliberately — "It is type-erased on purpose… No generics
are involved, and none are needed" — and monomorphisation is what would buy it back, at the cost the spec declined.

## Unions, and the tag they carry

`defunion` parsed and its shape was checked long before this; naming the type (`check.ml:312`) and constructing a
value (`:1075`) were both refused as milestone 6. They are not any more.

### It is closer than the milestone number suggested, and the reason is `Option`

`Tast.arm` already carried `acase` (a case name) and `binds` (the slots a payload binds to). That is union shape,
built and exercised, because **`Option` is a two-case union wearing a special coat** and `match` over it worked
already. So `check_match` grew a second *subject* rather than a second path: one function decides which case each arm
names and what type each name it binds has, and everything after it — the dead-set join across alternatives, the
result type, the exhaustiveness check — is the code that was there.

`Option` was **not** desugared into a declared union, and that is deliberate: `Option` is generic and no declared
union is. The coat is the part that cannot be taken off until generics exist.

A union is `Types.Named` exactly as a struct is. One case in `Types.t` covers both, and *which table the name is in*
is the only thing that tells them apart. That is what let a union be a field, a parameter, a return type, a slot and a
copy without a single one of those paths learning that unions exist.

### The layout, which is the part that has to be exactly right

```
%"U"       = type { i32, [k x iA] }
%"U.Case"  = type { the case's fields, in declaration order }
```

A tag, then room for the largest case, with `A` the alignment the *widest member of any case* needs and `k` the size
rounded up to it. Writing the payload as an array of `iA` rather than of `i8` is what makes LLVM align it without an
explicit `align` on a type — and it is what makes the whole thing

```c
struct { int tag; union { ... } u; }
```

byte for byte. **That agreement is the point.** A macro is `[Form] -> Form`, so `Form` has to be the same bytes in
the compiler and in the `dlopen`ed macro, and nothing at run time would notice a disagreement.

The tag is an `i32` and not an `i8`. With an 8-byte payload alignment the two cost the same, and `i32` is what a C
`enum` field is — the spelling the macro lane will have to write by hand on the other side of the boundary.

Both the union and one struct per case are emitted as *named* LLVM types, so every reader geps rather than computing
byte offsets of its own. Construction is a `store` of the tag and a `store` of the case struct through a gep into the
payload; `match` is a `load` of the tag and a gep the other way. There is one function that knows how a payload is
read — `case_field_addr` — because there are two readers: a `match` arm's binds and the structural printer.

The layout goes through the same oracle the DWARF section uses: LLVM's own answer for the emitted type, read back as a
constant-folded `ptrtoint`. Two unions are checked, one whose widest case is a pair of `f64` and one whose cases are
all `i32`, so the payload size and alignment cannot be constants the test agreed with by accident. The `Shape` in
`test/programs/unions.flan` was checked against clang's answer for the same declaration in C: 32 bytes aligned 8 with
the payload at offset 8, and 40/8 for a struct holding one.

### The surface

```clojure
(defunion Shape
  [Empty
   (Dot  [x f64  y f64])
   (Rect [w i32  h i32])
   (Tag  [name string  n u8])])

(Shape.Rect {.w 3 .h 3})     ; a value: the type and the case, then the fields
(Shape.Rect {.w 3})          ; ZII, exactly as in a struct literal — .h is 0
Shape.Empty                  ; a case with no fields is a whole value, not a call

(match s
  Empty        "empty"
  (Dot x y)    "a dot"       ; positional, in declaration order
  (Rect w h)   "a rect"
  (Tag name n) name)
```

**Construction is qualified and a pattern is bare.** The two are not inconsistent: at a constructor nothing says which
union is meant, and at a pattern the scrutinee's type already does. The qualified spelling is accepted in a pattern
too, since that is how the value was written and writing it again should not be an error. Two unions may therefore
share a case name, and that is not refused — refusing it would be a restriction with no mechanism behind it.

Construction needed **no change to `parse.ml`**. `.` is a symbol constituent, so `Shape.Rect` reads as one name; the
struct-literal arm fires on a symbol followed by a map; and the field-access arm needs a *leading* dot, so the two
cannot collide. `(Name {.field value})` is one syntax for both, and which it is, is decided against the tables.

### Tags are declaration order, so case order is part of the contract

Tag *n* is the *n*th declared case, from zero. The consequence is that an all-bytes-zero union is **the first declared
case with a zeroed payload** — which is exactly the rule that makes an `Option`'s zero a `None`, and it is what makes
`(defvar u U)` and an omitted union-typed struct field mean something rather than nothing. Reordering a union's cases
is a layout change, the same way reordering a struct's fields is.

### What is refused, and why each

- **A non-exhaustive `match`.** Refused, not defaulted, and the message names the cases with no arm. A match that fell
  through would have to produce a value of the match's type out of nothing, and the case a union grows tomorrow is
  exactly the one a reader wants to be told about today. `_` is how to say "the rest", written where it can be seen.
- **A case pattern binding some of a case's fields.** All of them or none, positionally — binding a prefix reads the
  wrong field the moment one is inserted above it.
- **Two arms for one case**, which is a mistake and never an intent.
- **A union with no cases.** No value of it can exist, so a parameter of that type is a function nothing can call.
- **A case field that is move-only**, in the same words a struct field already gets and for the same reason:
  recursive teardown arrives with `drop`.
- **A union as a map key.** The payload past the case in hand is indeterminate, so hashing the blob would make two
  equal values hash differently. Hashing one properly is a per-case walk driven by a switch — a different shape from
  the field list `struct_key_pair` emits, and nothing has wanted it.
- **`uninit` on a union.** This is the one refusal that is *not* the struct rule. Everywhere else `uninit` is an
  opt-out from ZII and the bytes are whatever they were: a garbage `f64` is a garbage number. A union's tag steers
  control flow — a tag no case names falls past every comparison in a `match` into a block LLVM is entitled to assume
  cannot be reached. That is the one place where garbage becomes "the optimiser may do anything".
- **`(.x u)`.** A union's fields belong to a case, and which case is being held is what the tag says. `match` is how
  one is opened, and its arms bind the fields of the case they matched.
- **A global initialised with a case.** `(defvar g U (U.B {.x 1}))` would mean serialising the fields into the payload
  blob at link time, which is a byte-level encoder this compiler does not have and which could not express a `string`
  field at all — that is a pointer the linker has to relocate and a byte array has nowhere to put a relocation. A
  *zeroed* global is fine and needs none of it: it is the first declared case.

### Recursion, and the shape `Form` will have

A union that contains itself by value has no finite size, and `payload_lay`
would recurse forever laying one out rather than failing. It does not get the
chance: `check_finite` already walked a union's cases, so `(defunion T [Leaf
(Node [l T r T])])` is refused with *"T contains itself by value, so it has no
size — go through (Ptr T)"*, and so is a pair of unions that contain each
other. Through a pointer it works, and that is the shape a `Form` has:

```clojure
(defunion Tree [Leaf (Node [l (Ptr Tree)  n i32])])

(defn depth [t (Ptr Tree)] i32
  (match (deref t)
    Leaf 0
    (Node l n) (+ n (depth l))))
```

### Why `match`'s fall-through is still `unreachable`

The block after the last case comparison is `unreachable`, kept from the
`Option` path. That is only sound if no reachable program can hold a tag no
case names — and none can: `Zero` is tag 0, which is a real case; every
construction writes a tag the checker resolved; and `uninit`, the one way to
get bytes nobody wrote, is refused on a union for exactly this reason. The
refusal is what pays for the `unreachable`.

### The diagnostics bug, fixed

`(A {.x 1})` on a case of a union said **"unknown struct A"**, because `env` had no table of case names and could not
tell a case from a misspelling. It has one now, keyed both by the full spelling `U.C` and by the bare `C`. The full
spelling is a *key* rather than something split out of a dotted name at the use site, because a union's own name can
contain a slash — an imported `rl/U` — and string surgery would own an edge this does not have to.

### Printing

`render.ml` walks a concrete type for `print`, the REPL inspector and the break buffer's locals, and a union fell
through its `Named` arm to `<Shape>`. It recovers the case from the tag by a chain of comparisons — the same shape the
enum arm already had, and for the same reason: the name is erased before any backend sees it — and reads the fields of
**that case only**. It prints `(Shape.Dot {.x 1.5 .y -2.5})`, which is what the source would write.

### What is left

An **imported** union is still refused by name at `load.ml:312`, so a union is file-local. That is not a blocker for
the macro expander: the prelude is parsed and prepended into the same flat namespace before `collect` runs, so a
`defunion Form` in `prelude.ml` is an ordinary same-file declaration needing no import — verified by declaring one
there and matching it from a program. `dev.ml`'s inspector still says "union values are milestone 6" for a stopped
frame's locals, and `shim.ml`'s "a Flan union has no C layout" is now inaccurate as prose though the refusal it guards
is still right: a union has a C layout and still may not cross to C by value, because the shim flattens aggregates.

## `defer` may be written in a `let`

The whole of this project's resource-cleanup answer, and NEXT.md records `drop` and a `with-cleanup` form as both
considered and rejected before it.

`defer` is a **compile-time** construct: the cleanup is copied into every exit path of the function. That is why a loop
body and a branch stay refused — a loop body's would fire once at function exit rather than once per iteration, and a
branch would have to express "maybe registered", which a form copied into every exit path or into none cannot say.

A `let` is neither. It is not a frame here: its bindings are function slots like any other and **nothing is released at
scope exit**, so a `let` at the top level of a function body has exactly the function's extent and a `defer` written in
it always registers. It was refused for a reason that does not apply to it.

The rule, stated precisely, because "top level of a function body" is easy to get wrong: a form at the top level of the
body may carry one, and so may a form in the body of a `let` that itself may — to any depth. A `let` inside a `while`
or an `if` has the loop's extent or the arm's, and inherits the refusal rather than the permission.

The implementation detail worth keeping: **the permission is granted again before every form of a body, never once
around the body.** `check` withdraws it as it starts, so granting it once would let the first `defer` through and refuse
the second — and two resources acquired in one `let` is the case this exists for. `defer-let.flan` covers exactly that,
along with nesting, interleaved registration order across the `let` boundary, and an early return.

This **amends `spec-memory.md`**, which states under "When storage is released" that `(defer (free v))` for a
`let`-bound `v` is "not expressible today" and that no idiom in the spec may depend on it. It is expressible now.

`do` at the top level of a body has the same extent argument and is deliberately **not** included: nothing asked for it,
and the rule is easier to state and to trust with one construct in it.

## Assets are baked in, and the reason it is a compiler feature

NEXT.md decision 1. `(embed "brush.png")` is a `[u8]`, `(embed "brush.png" string)` is a `string`, and
`(embed-dir "assets")` is a `[n EmbedFile]` sorted by name. Odin's `#load` and `#load_directory` are the model
(`src/parser.cpp`, and `check_load_directive` / `check_load_directory_directive` in `src/check_builtin.cpp`); Odin's
`#` is not imported, because an s-expression language already has a head position for a name and these resolve as
ordinary named calls exactly the way `vec-new` and `heap-allocator` do.

**The reason for this shape rather than a build flag is the one that decided it.** It is a *compiler* feature, so it
needs no linker arguments and no per-target packaging, and it works identically on desktop and web. That matters more
here than it does for Odin, because `Load` hands out `lflags` only to a directory package and `main` is not exported,
so a program can never be a package: the single file doing `(rl/load-texture "brush.png")` is structurally the one file
with **no link channel at all**. The web lane found that hole and did not invent a flag for it. Embedding has no such
hole, because there is nothing to tell the linker.

**It costs nothing at run time.** The bytes reach the program as a `Tast.Str` node typed `[u8]`, which emit.ml turns
into the same `private unnamed_addr constant` every string literal already becomes, and its `escape` is byte-exact
across the whole 0–255 range, so a PNG survives the round trip through the `.ll`. Bound with `defconst` at top level an
`embed-dir` is an LLVM constant outright, through emit.ml's `const`.

**A `Str` node typed `[u8]`, not a `Bytes` prim over a `string`.** This is the one non-obvious choice. `Bytes` is
identity — emit.ml lowers `Types.String` and `Types.Slice _` to the same `%slice` — but wrapping the literal in a prim
makes the node non-constant, and `const` then refuses an `embed-dir` in a `defconst` with *a global's value must be a
compile-time constant*. Both of emit.ml's string emitters take the bytes and ignore the node's type, so it is the same
constant either way and this one is a constant a global can hold.

**Two spellings, not one form that changes type with its context.** Odin threads a `type_hint` everywhere and can
afford `#load("p")` to mean a `string` here and a `[]u8` there. With structural equality, no implicit widening and no
coercion anywhere, the same text meaning two types would be a wart, so `string` is written down when it is wanted. The
site's expectation is a fallback only and nothing depends on it.

**The path is a literal and resolves relative to the file the form is written in.** Both are Odin's rules and for
Odin's reasons: the bytes must be in hand before any value exists, which is what makes the result free; and a path
relative to the compiler's working directory would make a package's assets depend on where `flan` was invoked from,
which cannot be right. A missing file is a compile error naming it, never an empty embed — an asset silently absent is
exactly the quiet wrongness this removes. An empty *directory* is not that case and embeds cleanly as `[0 EmbedFile]`.

**The directory lookup is a linear scan, and that is the chosen answer rather than the fallback one.** `embed-find` is
an ordinary prelude function over a `[EmbedFile]`. A directory embed is tens of entries whose names sit in cache-warm
`.rodata`; a compile-time perfect hash would be a build-time map with its own failure modes that nothing has asked for,
and sort-and-bisect is the next step if a program ever embeds thousands of files — it would not change the type. It
takes a **slice** rather than the array, because an array's length is part of its type and there are no generics, so
the call reads `(embed-find (slice assets 0 (len assets)) "brush.png")`. Entries are sorted by name because `readdir`
order is filesystem-dependent and an unsorted embed would make two builds of identical sources emit different `.ll`.
Non-recursive, files only — Odin again.

**The sharp edge, inherited and not widened.** The slice points into `.rodata`, so a store through it segfaults at
`-O0` and is deleted as undefined behaviour at `-O2` — the same trap the prelude's ASCII-case note measures for
`(bytes "Hi")`, and the same one NEXT.md tracks as "writing through a string literal". Nothing here makes it worse and
nothing here fixes it; provenance is what would. **To get a mutable copy, clone the bytes into a `Vec`.** It is worth
saying loudly because an embedded asset is precisely the thing someone will try to decode in place.

**What this does not do.** `sand.flan` still calls `(rl/load-texture "brush.png")`, which hands raylib a path for
raylib to open. Pointing raylib at embedded bytes needs `LoadImageFromMemory` and `LoadTextureFromImage` in place of
`LoadTexture` — a raylib binding question, not an embedding one — so the flagship program is not yet asset-free on the
web. The mechanism it needs is in.

## `slurp`, `barf`, and the two ways they fail

NEXT.md decisions 2 and 5. `(slurp path)` and `(slurp path allocator)` read a whole file into a `(Vec u8)`;
`(barf path bytes)` writes one.

**`slurp` waited for `Vec` because its result has no length until the file is read**, and it obeys spec-memory.md's
rule without an exception: *no allocating operation returns an error*. There is no `Result` here, no out-parameter and
no error code — `slurp`'s type is `(Vec u8)` and `barf`'s is `Unit`.

**Two failures, two conditions, and the guards nest rather than merge.** Allocation failure is `StorageExhausted` under
`retry`, unchanged and reused. File failure is `FileError {.path .op .reason}` under `retry` and `use-value`. They stay
apart because they ask two different answerable questions: the handler that grows an arena is not the handler that
supplies another path, and collapsing them would make one handler guess which it was looking at. `file_guard` in
check.ml is `alloc_guard`'s shape built from the same nodes — a `while`, a `restart-case` and an `error` — so the
backend learns nothing new.

**The restarts are Common Lisp's pair for a `file-error`.** `retry` for "the file may be there now"; `use-value [p
string]` for "try this other path". `use-value` is the first restart clause the **compiler itself** emits with a
parameter — typed restarts landed the same session — and its parameter *is* the path slot the attempt reads, so the
clause body is empty. emit.ml's `bind_params` stores the invoker's argument into the slot, the clause falls through,
and the loop re-attempts against the new path. Everything is inside that loop, so a `use-value` naming a different file
re-measures it and re-allocates for *its* size; the `Vec` is freed at the top of each turn, which is why a retry does
not leak, and freeing a `Vec` that never allocated is a no-op.

**The two forms resolve paths by opposite rules, and it is worth saying in one place.** An `embed` path is resolved at
*compile* time relative to the file the form is written in. A `slurp` or `barf` path is resolved at *run* time by the
host, against the process's working directory — these are ordinary values, and one can arrive from `argv` or from a
`use-value` restart. `test/programs/slurp.flan` reads `"programs/assets/a.txt"` only because the suite runs from
`_build/default/test`. Two forms in one section with opposite rules is exactly where someone gets bitten.

**The break loop can be offered this restart and cannot fill it in.** That is not new behaviour, only a new way to
reach it: a break loop chooses a restart by position and has nothing to supply a parameter with, and emit.ml already
emits a `flan_restart_unarmed` guard on every clause that takes one, so taking it refuses with the reason rather than
running the clause on a zeroed buffer. `slurp`'s `use-value` is simply the first such clause the *compiler* emits —
`alloc_guard`'s `retry` takes no parameters — and it rides the same path a hand-written one does. The acceptance suite
asserts the guard is on the emitted IR for both.

**On the web, `barf` signals — every time, with the path in the condition.** This is the decision worth restating,
because two more obvious answers are both wrong here. A **build-time refusal** is unusable: Flan has *no conditional
compilation*, nothing in `parse.ml` or `check.ml` reads the target, so "isolate this to desktop" is not expressible in
source and the refusal would have nowhere to be silenced. A **silent no-op** is worse than either, because that is how
a save file disappears with nothing said. So the program gets a condition and decides — which is the language having
something Odin does not. Odin stubs its whole file API on js/wasm to `.Unsupported` (`core/os/file_js.odin`) so that
importing `core:os` "panics cleanly", and a panic is not a decision. **The restriction was taken; the mechanism was
not.**

Nothing in the compiler reads the target to do this. The refusal is one `#ifdef __EMSCRIPTEN__` in `flan_rt.c`, which
is where the host ABI is *already* implemented twice. `slurp` keeps working on the web — it compiles, runs, and reports
a missing file honestly — and the bytes a web program actually wants come from an `embed`.

`test/programs/web-files.flan` is one source built for both targets, and `test/test_web.ml` **runs** it under node
rather than asserting the artifact's shape: an artifact-shape assertion would say nothing about the thing the decision
bought. The test checks that the refusal is printed with its path and reason, and that the desktop's success line is
*absent* — a silent no-op would have taken that branch.

### What the host ABI grew by, and why that much

plan.org names the filesystem as the #1 portability risk — "pack assets, one abstraction, never touch paths" — so the
widening is written down rather than assumed. It is **three calls and one reader**:

| Call | What it does |
|------|--------------|
| `flan_file_size(path, n, &out)` | how many bytes are there |
| `flan_file_read(path, n, buf, cap, &got)` | fill a buffer the caller owns |
| `flan_file_write(path, n, buf, len)` | write a whole file |
| `flan_file_fail_reason()` | which of the four reasons it was |

They are POSIX-shaped and **Vec-ignorant**: no handle crosses the boundary, nothing is held between calls, and each
takes a path and answers 1/0 the way every allocator entry point already does. `flan_slurp_into` — the part that knows
what a `Vec` is — is runtime *glue* on this side of the ABI, not a fourth host call, so a second target implements
three functions and inherits the rest. The reason is a global rather than an out-parameter for the same reason
`flan_alloc_fail_bytes` is: the condition is a value struct on the failing frame's stack with fixed numeric fields and
no rendered message.

**These do touch paths, which is the widening plan.org warned about**, and decision 2 took it knowingly. `embed` is the
half that does not: it needs no host ABI at all, so "pack assets" remains the answer for anything known at build time
and `slurp` is for bytes that genuinely are not.

`FileError` is one type with a `reason` field rather than a family, because conditions have no hierarchy today
(spec-conditions.md §1) and a family would need one handler clause per member to say "any file error". NEXT.md decision
4's parent link is the answer to that and is not built; when it is, these reasons can become types without any call
site changing.

**One thing `slurp` and `barf` reveal that is not theirs to fix.** A handler that wants "try to save, and carry on if
you cannot" has nowhere to go: `error` is diverging (spec-conditions.md §2), a handler returning normally has not
answered it, and neither `retry` nor `use-value` means *give up*. `web-files.flan` calls `exit` for that reason. A
`continue`-style restart — or `handler-case`, which unwinds — is what the case wants, and neither exists.

## Where build time goes

`flan build calc-me.flan` was ~160ms, and ~95% of it was clang. **The object cache is in**, and it is now ~110ms:

| Step | Cost |
|---|---|
| frontend: read → parse → load → check → emit | <10ms, below the timer |
| `clang` on the `.ll` | 60ms — `llc` does the same codegen in **20ms** |
| `clang` on `flan_rt.c` | 40ms — **now cached, paid once** |
| link | 20ms |

Every C translation unit a build needs — the host shim and each package's shim — goes through `Build.compile_c`, which
compiles to a `.o` under `$TMPDIR/flan-objcache` and reuses it. The key is a digest of the source text, the compiler
(its path, size and mtime, so an upgrade invalidates without paying a `clang --version` subprocess per build),
`opts.opt` and `opts.target`. The opt level has to be in there: the acceptance table builds the same programs at `-O0`
and `-O2`, and an `-O2` object must not serve an `-O0` build. The object is written to a temporary name and `rename`d
into place, so two concurrent builds cannot see a half-written one.

Measured: calc-me 160ms → 110ms; sand ~720ms → ~700ms, since sand's time is mostly linking libraylib and its C was never
the cost. The cache is keyed by content, so it never needs invalidating by hand — `rm -rf` on the directory is only ever
a disk-space decision.

The other cheap win is still open: skip the clang driver for the `.ll` (`llc` + link directly), worth another ~40ms. It
is a subset of the dev path's machinery. Check `llc`'s major version against clang's before relying on it — the emitted
IR text is currently absorbed by the driver behind `-Wno-override-module`, and a version mismatch surfaces as IR parse
errors.


## The order it was built in, and why that order

Every item here is done; it is kept because the *sequencing* is the part worth remembering. Decided in conversation:
wasm32 could wait (believed to be a solved problem once the builtins archive was in place), and **the dev loop is the
thesis of the project**, so it came first. Staged so each step was runnable on its own — the failure mode being to
build a daemon and a protocol before knowing whether the reload primitive worked.

1. ~~**The reload primitive, measured.**~~ **Done** — `Emit.redefinition`, `Build.shared`, `test/reload_host.c`, ~19ms.
See the section above.
2. ~~**Indirection cells.**~~ **Done** — `Build.opts.dev` / `flan build --dev`, `flan_reload_install`,
`runtime/flan_dev.c` for names introduced at run time, and a fixture where an untouched call site follows the swap and a
run-time-added function is itself redefined. See the section above.
3. ~~**The agent, in C.**~~ **Done** — `vendor/agent/`, a listener thread that loads and a game thread that installs,
and sand.flan polling at the top of its frame. See the section above.
4. ~~**The daemon**~~ and ~~**5. the Emacs client**~~ — **both done**, and the protocol is s-expressions rather than
nREPL's bencode; see the two sections above for why that changed. An nREPL front end can sit on the same `Session` if
something else ever needs to talk to it.

**One decision left to settle before step 2**, because both change codegen and are painful to retrofit:

- ~~**Do cells cover globals, or only functions?**~~ **Settled by step 1: functions only.** A redefinition module
declares every global `external`, so globals live in the host and survive a reload — which is what "edit the code, keep
the sand" needs. The consequence to watch is the other half: adding a `defvar` to a file cannot take effect on reload,
and changing one's type is a silent mismatch against storage the host already laid out. Nothing detects that yet.
- **What is a redefinition unit — one function, or a file?** A file is much easier to make correct and is what
`load-file` wants anyway; one function is what `C-c C-c` wants and is where the 16ms number comes from.

Deferred until after the dev loop:

6. ~~**wasm32.**~~ **Done, with one glued joint.** `flan build --target=wasm32-wasi` produces a module, and
`test/programs/sand-headless.flan` prints `15595743031174623232` under it — the same hash as native, byte for byte, at
`-O2` and at `-O0`. That is the milestone: the RNG is ours rather than libc's precisely so that number can be compared
across targets, and it compares equal. `values.flan` and `machine.flan` run there too, which is where a 32-bit pointer
would have shown. The acceptance table runs all four, and skips them by *probing* — it builds the smallest program and
runs it — rather than by looking for a binary on PATH.

Three things this cost that were not in the old note:

- **The entry point is not `main`.** wasi-libc's start code calls `__main_argc_argv`; clang renames C's argc/argv `main`
to that, and the `.ll` `Emit` writes says `@main` literally. The link succeeds and the program traps on a
signature-mismatched weak stub. `Build.wasm_main_source` is a two-line C shim that bridges it, and the `__asm__("main")`
label in it is load-bearing: spelling the callee `main` makes clang rename *that* too and the shim becomes an infinite
self-call.
- **The target has to reach the C compiles, not just the link.** `flan_rt.c` includes `<stdio.h>`; without `--sysroot`
it never gets that far. `target_flags` is computed once and passed to both, and the whole flag list — not just the
triple — is in the object cache key, so repointing a sysroot cannot serve a stale `.o`.
- **Fedora's sysroot is one level deeper** than wasi-sdk's: `include/wasm32-wasi/stdio.h`, not `include/stdio.h`. Both
shapes count.

**The glued joint, and the one thing this contradicts in the old note.** The old note said the builtins archive has to
come from a wasi-sdk release. It does not have to: emscripten builds the same compiler-rt for wasm32 and calls it
`libcompiler_rt.a`, and dropping that in as `libclang_rt.builtins.a` links and runs. It is a different triple
(`wasm32-unknown-emscripten`) built by a different clang (22 against Fedora's 20), so it is *substituting*, and wasi-sdk
is still the proper article. `build.ml` looks for `FLAN_WASM_BUILTINS`, then `/opt/wasi-sdk/...`, then emscripten's
beside `emcc` on PATH, and refuses by name listing every path it tried when none is there. clang's resource directory is
root-owned, so the archive is not dropped into it — a shadow resource directory is built under the object cache, named
by a digest of clang's own resource dir plus the archive's path, size and mtime, with the real `include` symlinked in.

**The runtime is Node.** No `wasmtime` and no `wasmer` on this machine; `test/wasm-run.mjs` is twenty lines of
`node:wasi` and the table prefers `wasmtime` or `wasmer` if either appears. `--no-warnings`, because `node:wasi` writes
an `ExperimentalWarning` to stderr on every run and the harness compares combined output.

**Refused by name, not half-supported:** `--dev` with a wasm target (the reload path is `dlopen`), `Build.shared` with
one (same reason), and `flan run --target=` (a `.wasm` is not something this host execs — build it and point a runtime
at it).

~~Still open: raylib on wasm, which plan.org wants through emscripten and its own sysroot.~~ **Done — and it is a third
target, not a mode of this one.** wasi-sdk is right for the headless table and was never going to be right for the game
build. See the next section.

## The browser is the third target

`flan build --target=web` produces a page, its JS and a `.wasm`, and a raylib example opens in a browser from source
that was not touched. The two wasm targets share the word and almost nothing else, so `is_wasi` and `is_web` are
separate predicates and `is_wasm` is their union — the union is exactly the set of facts about the *machine* (32-bit
pointers, no `dlopen`), which is what the refusals are about, and nothing else is shared.

**The compiler is `emcc`, not `clang`, and that is the whole of the sysroot story.** Everything the wasi target has to
find by hand — a sysroot, a builtins archive, a shadow resource directory, the `__main_argc_argv` shim — is what emcc
*is*. `target_flags` for `web` is the empty list; the only thing checked is that emcc exists, refused by name where the
reason can say so. The one fact that had to be true for any of this: **emcc takes a `.ll` on its command line**, which
it does, so `Emit`'s output needs no change and the IR stays target-independent. The object cache keys on the compiler
binary's path, size and mtime as it always did — now of *whichever* compiler the target uses, so an emcc `flan_rt.o`
and a clang one cannot collide.

**The main loop: `-sASYNCIFY`, not `emscripten_set_main_loop`.** The older note above had the browser fact right —
it cannot be blocked — and drew the wrong conclusion from it. `emscripten_set_main_loop` wants the loop body as a
callback, so every one of the eleven examples that writes

```lisp
(until (rl/window-should-close?) ...)
```

would have to be split by hand into an init and a tick, and the web program would stop being the native program.
Asyncify rewrites the module so a call can suspend across a return to the event loop, and raylib's web platform is
built for precisely that: `WindowShouldClose()` on `PLATFORM_WEB` is an `emscripten_sleep(16)` that then returns false
(raylib 5.5, `platforms/rcore_web.c`, read rather than assumed). So the loop yields once a frame at a call it already
makes, and **no example changed a character**. The price is real and is paid by every web build: asyncify instruments
the whole module, roughly doubling code size. It is not applied per-program because "does this program block" is not a
question `Build` can answer, and a flag set that varies per program is a cache key that varies per program.

**`link` lines can be addressed to a target.** `vendor/raylib/link` named `libraylib.so.550`, which exists on the host
and nowhere else. A line may now carry `@native`, `@wasi` or `@web`, an untagged line applies everywhere — which is
what every existing `link` file already is — and `${NAME}` expands from the environment. The selection happens in
`Build` and not in `Load`, which is where the file is read, because **`Load` resolves imports before a target is
chosen**: the same program is built for both, and a package's linker arguments arrive here as a flat list of strings.
`Load`'s part in this is to pass the lines through untouched, which it already did.

**raylib for the browser is built, not installed.** No emscripten port provides it (`emcc --show-ports`: there is
`contrib.glfw3` and no raylib), so `vendor/raylib/build-web.sh` clones raylib at the **5.5** tag — the one whose
`.so.550` the host links, because `raylib.flan` carries raylib's struct layouts and enum values and two targets built
from different raylibs would disagree about them in silence — and compiles the seven modules with
`-DPLATFORM_WEB -DGRAPHICS_API_OPENGL_ES2` into one archive under `vendor/raylib/web/` (gitignored). `rglfw.c` is not
among them: the web platform uses emscripten's own GLFW port, which is why `link` carries `@web -sUSE_GLFW=3`. No
headers are installed, for the reason the host build needs none — the generated shim declares its own prototypes.

**The HTML shell is a string in `Build`, not a file in the tree**, for the same reason `Runtime_src` is: it has to be
wherever the compiler is, and a build that cannot find its own shell fails for a reason nobody spelled. It is a canvas
and a `Module.print` that puts stdout on the page; `FLAN_WEB_SHELL` replaces it. `--shell-file` is passed only when the
output is a `.html`, because emcc accepts and ignores it otherwise.

**Refused by name, inherited whole:** `--dev`, `--debug`, `Build.shared` and `flan run --target=` are refused for
`web` exactly as for `wasm32`, each naming `web` rather than `wasm32` in the message. `--sanitize` is refused too, but
the web half of that refusal is weaker than the wasi half and says so: emscripten *does* ship an ASan, and nothing here
has ever run it. A sanitizer that has never been run is one whose silence means nothing.

**What the test can honestly check.** `test/test_web.ml` is headless and permanently so. It probes — emscripten may not
be installed, and the raylib archive is not in the tree — and skips with the reason rather than going red. What it
asserts: the three files exist, the module starts with `\0asm`, the page references its own JS and carries the canvas,
and node runs the emitted JS and gets `ok`. For raylib it builds `core-basic-window.flan` unchanged and then reads the
module for the two things that would be false if the mechanism were wrong: an `asyncify_start_unwind` export, and a
`glViewport` import that can only have come from raylib's web platform. Import and export names are plain strings in
the binary, so this needs no wasm reader.

## sand.flan in a browser

**The flagship program builds for the browser and the artifact opens.** Three things were between it and the target,
and none of them was the `#include` the earlier note named.

**The brush was a path, and a path is what cannot work.** `(rl/load-texture "brush.png")` hands raylib a filename to
open; a bare relative path has no meaning where there is no filesystem, so raylib would have opened nothing and the
cursor would simply have been missing. It is `(embed "brush.png")` now, decoded by a new binding —
`LoadImageFromMemory`, which completes the chain embedded bytes -> `Image` -> `Texture2D` that `LoadTextureFromImage`
already had the other half of. The declaration is `(Ptr u8)` plus an explicit `i32` count, because `shim.ml` refuses a
slice parameter and its refusal says exactly that; the Flan wrapper beside it takes the slice apart, which is the
idiom `collision-point-poly?` and `load-font-ex` already established. One decode serves both textures now: the
unflipped upload first, then `ImageFlipHorizontal` in place, then the mirrored upload — **the order is load-bearing**,
and if the two badges look the same it was swapped.

`load-texture` and `load-image` lose their only call site in this repository by this change. That is deliberate rather
than an accident of editing: a path-based load is the one shape the browser cannot have, and the bindings stay for the
desktop programs that will want them.

The embedded path resolves **relative to the file the form is written in**, which is why
`test/programs/sand-headless.flan` still works: it reaches `sand.flan` through `../../` out of a sandboxed `_build`,
and the PNG is found beside `sand.flan` and not beside the working directory. `test/dune` therefore lists `brush.png`
as a dependency of every stanza that builds sand — an embed is read by the *checker*, so it is a build input and not a
run-time one. The same fact reached `test_session`'s `C-c C-k` case, which re-evaluates sand.flan's whole text: it now
passes `~origin`, which is the buffer path both editor paths already send, because the default `<eval>` origin would
resolve the embed against the working directory instead.

**A package's C may be addressed to one target, the way a `link` line already could.** `Load` collects a package's
`.c` files by listing the directory, and there is nowhere in a directory listing to put a tag except the name, so the
tag goes there, before the extension:

```
vendor/agent/flan_agent.c        compiled everywhere, unless displaced
vendor/agent/flan_agent.web.c    compiled for the browser, and displaces the above
```

One rule: **a tagged file is compiled only on its own target, and there it replaces the untagged file of the same base
name.** Untagged is the default and every existing package is untagged, so nothing that did not opt in changed.
Replacement rather than the pure tagging Go's `_windows.go` and Odin's `file_js.odin` use, and the difference is the
point — pure tagging would mean renaming `flan_agent.c` to `flan_agent.native.c` to teach the package about a target
it had never heard of, and this way a package gains a target by gaining a file. The selection is in `Build` and not in
`Load`, for the reason `select_lflags` gives: `Load` resolves imports before a target is chosen.

**The dev agent on the web is a no-op, and it is not the `barf` decision being contradicted.** The compile error that
led here — `struct timeval` incomplete, because emscripten's headers do not pull `<sys/time.h>` in transitively — is
the surface. Underneath: *the agent is a socket server and a browser has no sockets*. Adding the include produces an
agent that compiles, links, starts and can never accept a connection.

Refusing `vendor:agent` on a web target was the other candidate and is ruled out by arithmetic, not taste. Flan has no
conditional compilation, `sand.flan` calls `(agent/start ...)` unconditionally, and `Reach` cannot prune a package
something reachable calls into — so a build-time refusal means the flagship program does not build for the browser at
all without being edited into a second program. **A refusal is only honest when the caller has a way to not ask.**

The two decisions look contradictory and are not, and the difference is *what the caller loses*. `barf` is asked to
make something durable; a no-op returns success to a program that now believes the bytes are on disk, and the loss is
real, is the user's, and is discovered later or never. The agent is asked to accept redefinitions from an editor; on
the web there is no editor, no socket and no session — `--dev` is refused by name on every wasm target, so a web build
has no cells to install a redefinition into even if one arrived. **Nothing is lost because there was never anything
there.** `sand.flan` already says the same about a *native* release build, at the call site: "Building without `--dev`
is fine — nothing has cells to install into, so a module is refused on the listener thread and the loop never
notices." A web build reaches that outcome by a shorter route. `start` returns `-1`, which is what `flan_agent.c`
returns for a path it cannot bind; `poll` and `wait` return 0, which is what the native build returns on every frame
nothing arrived on. The whole argument is written at the top of `vendor/agent/flan_agent.web.c`, where the next reader
will meet it.

### Building it and opening it

The raylib archive is built once and is not in the tree. From the repository root:

```sh
sh vendor/raylib/build-web.sh          # clones raylib 5.5 and compiles it with emcc
export FLAN_RAYLIB_WEB=$PWD/vendor/raylib/web/libraylib-5.5.a
```

`build-web.sh` prints that `export` line itself. Then:

```sh
flan build sand.flan --target=web -o sand.html
```

**The output must be named `.html`.** `--shell-file` is passed only when it is, because emcc accepts and ignores it
otherwise — so `-o sand` produces a module with no shell, no canvas, and a page that looks like it built fine and
paints nothing. Three files land beside it, in whatever directory `-o` names: `sand.html`, `sand.js`, `sand.wasm`.

**A `file://` URL will not work.** The page fetches `sand.wasm`, and a browser refuses that from the filesystem. Serve
the directory holding the three files:

```sh
python3 -m http.server 8000
```

and open **`http://localhost:8000/sand.html`**. Left mouse paints; the keys are the ones the native build has.

### What only a human opening it can settle

Verified headlessly, by `test/test_web.ml`: the three files exist, the module carries `asyncify_start_unwind` and
`glViewport`, and **brush.png's own bytes are in the module, whole** — the assertion that keeps the embed from
rotting. Not `IHDR`: stb_image, linked in from raylib, carries that string itself, so an `IHDR` check would pass on a
build where the embed emitted nothing.

Not verified, and not verifiable here. `node sand.js` instantiates the module, runs `main`, and dies inside `glfwInit`
on `window is not defined` — which says the module is live and says nothing about the canvas.

- **Whether it paints at all.** Nothing in CI has ever seen a pixel of this.
- **Audio.** `start-audio` generates a tone, `ExportWave`s it to `/tmp/flan-sand-tone.wav` and loads it back as a
  music stream. That is raylib's own `fopen`, not Flan's `barf`, so the `#ifdef` in `flan_rt.c` does not cover it and
  emscripten's MEMFS may well give it a writable `/tmp`. Every use is behind `music-ok`/`tone-ok`, so a failure is
  silence and not a crash. Separately, browsers suspend the audio context until a user gesture, so `audio-ok` may be
  true while nothing is heard until the first click.
- **The loop never exits.** `WindowShouldClose()` on `PLATFORM_WEB` is an `emscripten_sleep` that returns false, so
  `until` never terminates and **none of `main`'s `defer`s ever run** — no `CloseWindow`, no `UnloadTexture`. That is
  correct for a page, which is torn down by the tab closing, and it is worth knowing before reading anything into it.
- **Canvas size against `screen-width`/`screen-height`.** The shell is a string in `Build` and its canvas is not sized
  from the program, so 900x600 may be letterboxed or cropped.

## The colon belongs to keys, so a field label is a dot

`{.x 1.0 .y 2.0}` is how a struct is constructed, and `{inner .field}` is how a pattern names one. The colon is gone
from both, and what is left of it is one job: keys — map keys and enum members.

**What was wrong with the colon.** Nothing, taken alone. The problem was that it had two jobs and the dot had one.
`(.x v)` already read a field; `{:x 1.0}` also named a field, while `:space` named an enum member. So the dot meant
"field" and the colon meant "field, or member, depending". Moving the label to the dot leaves each mark with one
meaning, and it costs nothing to read because **the delimiter already disambiguates**: `(.x v)` is a list and
therefore a call and therefore an access; `{.x 1.0}` is a brace form and therefore a construction. There is no
position where the two could be confused, which is why the same spelling can serve both.

**Why it had to land before `Map`, and this is the real reason.** A map literal wants to be `{:key value}`. While
struct construction owned that exact spelling, a map literal and a struct literal were *the same syntax*, and the
only thing that could tell them apart was what the checker expected at that position. That is a context-sensitive
grammar for no gain. Reserving the colon for keys keeps the two visibly distinct at the reader, before any type is
known. Doing it after `Map` landed would have meant changing both; doing it first meant changing one.

**`:keys` kept its colon, and that is the point rather than an inconsistency.** `{:keys [x y]}` is the one thing in a
brace that is not a field name — it is an instruction to the compiler that happens to sit there, and it takes a
vector rather than a value. Giving it a dot would have made the dot mean "a field, or the word keys". Leaving it a
colon lets the dot mean exactly one thing, *this names a field*, which is the whole reason the colon was given up.
Everything else Clojure puts in that position — `:as`, `:or`, `:strs`, `:syms` — is still refused by its own name.

**The old spelling is refused, not accepted quietly**, and the refusal names the new one: `a field label is written
.x, not :x — the colon is for keys`. Two accepted spellings is how two spellings become permanent, and the standing
rule here is that what is not supported is rejected explicitly with the reason. Both refusals are tested by their
reason, on the construction side and on the destructuring side, which is what stops the colon drifting back.

**The sweep is a tool, not a one-off.** `tools/colon-to-dot.py` converted 681 labels across 45 `.flan` files,
`vendor/` included, and 94 more in the Flan embedded in `lib/prelude.ml` and the tests. It works on *forms*, not on text: a keyword
becomes a dot only where it sits in a field-label position inside a brace, so an enum member in value position
(`{.k :hi}`), a genuine EDN map inside a string (`test/programs/edn.flan`), and a type-position `{K V}` are all left
alone. It was kept in the tree because several lanes branched before it and their Flan needs the same pass at merge.

**What it deliberately did not change: the printed form.** `render.ml` still prints `(V {:x 1.5 :y 0})`. That string
is a wire format — `emacs/flan-inspect.el` parses it back and hard-codes the colon when it reads a field out — so
moving the printer alone would break struct inspection in the dev loop without breaking any test that says so. The
printer moves when its reader does, in the Emacs lane. It is the one place the old spelling is still correct, and
the reason is worth keeping: **a format with two ends only changes at both.**