flan/DISCUSS.md

Let's discuss

Open questions, raised and deliberately not answered yet. Nothing here is a decision or a task. Each entry is the question as asked, plus what is already known in this repo that bears on it — so the investigation starts from what exists rather than from scratch.

Settled decisions live in NEXT.md. Reasons for what already exists live in BUILT.md.

---

1. i, the inspector, and the frame it cannot see — answered and built

Answered, and built as option 2. BUILT.md's "Two ways to root a walk, and why neither subsumes the other" is where it lives now, including the correction to what this entry said about rooting at an address. The number is kept because other files cite these by number; nothing here is open.

2. Annotating the IR and the disassembly with the source

The IR half is nearly free and should just be done. emit.ml writes .ll as text, so a comment costs nothing and cannot break anything, and every typed IR node already carries a Loc.t.

The disassembly half, with the optimisation question settled. objdump already interleaves source into a listing when DWARF is present (-S), and the daemon already shells out to objdump — so the -O0 case is close to a flag.

Settled in conversation: -O0 is expected to follow the source and gets the full annotation; -O2 is not expected to and gets either nothing or whatever best-effort mapping falls out. That removes what looked like the blocking question. --debug forcing -O0 is therefore fine and does not need decoupling for this.

How SBCL does it, since it came up. SBCL does not shell out. sb-disassem is its own disassembler, written in Lisp, that knows the instruction encodings directly. What that buys is annotation from the inside: it labels constants the function references, names the functions being called, marks entry points, and shows its own calling conventions — because it compiled the code object and still holds the metadata.

The relevant observation is that this project is closer to SBCL's position than the objdump route suggests. The daemon owns the build and holds the metadata too; it simply is not feeding much of it into the listing yet. Naming the function behind an indirection cell, or a constant by its source name, needs no instruction decoder — only the information the daemon already has. Writing a disassembler is not the interesting part and should stay off the table; richer annotation of objdump's output is cheap and is where SBCL's advantage actually comes from.

3. def, defvar, defconst — three roles, currently two words

There is no def. defvar is a mutable global that is always re-initialised on redefinition; defconst is a compile-time constant, folded into its use sites.

The problem in practice: C-c C-k re-evaluates every top-level form against a running program, so today it wipes state you may have spent a session accumulating.

Settled in conversation — three roles, three words:

• def — always re-initialised. The tweakable: some number you are adjusting and do not care about preserving.
• defvar — re-initialised only if it would come out different. For heavier initialisers: an N×M grid that should recompute only when the rows or columns actually changed.
• defconst — folded, and reserved for what genuinely cannot change for the life of the program. rows/cols in sand.flan are honest defconsts: the program does not support resizing on the fly and changing them means resetting everything anyway.

Note this defvar is not Common Lisp's, which is "assign only if unbound". This one is value-dependent, which suits a live-edited game better and is the harder of the two to implement.

The open question is what "different" means, and the grid example is what makes it sharp. If the grid is sized by rows and cols, the initialising expression's text is unchanged when rows changes — only its value is. So:

1. Compare the expression (textual or AST). Cheap, and does not catch the case above.
2. Compare the expression plus everything it depends on. Catches it. Needs dependency tracking.
3. Evaluate and compare the result. Correct, and defeats the entire purpose whenever the initialiser is expensive — which is the only case this feature exists for.

Option 2 is what is wanted.

The reason to do it: it is the same machinery defconst needs. A folded constant cannot be tuned live, because its value is baked into every function that used it — so the values you most want to tweak while the game runs are exactly the ones you cannot. Fixing that means knowing which functions depend on which constants, and rebuilding those. That is the same dependency tracking defvar's "if different" requires. Built once, it buys live-tunable constants and the conditional re-initialisation. Neither is worth building alone; together they are clearly worth it.

4. Structural typing, row polymorphism, anonymous structs

The goal as stated: make this feel Clojure-like while still being typed.

The representation is already structural. types.ml does structural equality on resolved types, and a Flan struct is exactly its C layout with no header and no tag word. What is nominal is the checking, not the data — so this is a front-end question, not a data-model change.

What it buys, in rough order of value here:

1. Your destructuring already looks like this, and that is the strongest argument. Clojure's {:keys [x y]} works in binding position today. Structural typing makes the same notation work in parameter position — the pattern you write to pull fields out becomes the signature saying what you accept. One notation, two places, no separate declaration. That is precisely the "Clojure-like but typed" feel being asked for, and it needs no new syntax, only a new meaning for syntax that exists.
2. Functions over "anything with these fields". Magnitude over anything with an x and a y — a position, a velocity, an enemy, a bullet. In a game this recurs constantly because everything has a position.
3. Ad-hoc returns. Returning a hit flag and a point without declaring a type for the pair. The Clojure habit of returning a map, typed and free.
4. The FFI stops needing conversions. raylib's Vector2 and a locally-defined vector are byte-identical and are today two nominal types. Structurally they are one.

Two costs, and the second is the real work:

• Not an "explicit over convenient" question — that framing was wrong and is recorded so it is not repeated. The standing rules about implicit numeric conversion and about a bare integer not becoming an enum are about lossy things happening silently. A structural match is not lossy; it is a different notion of type identity. The two were lumped together in conversation and the author rejected the pairing, correctly. Strong typing and nominal typing are separate choices and this only touches the second.
• Field order is layout, and that is the hard part. {x f32, y f32} and {y f32, x f32} have the same fields and different memory. So either field order becomes significant — surprising, since nothing else about a structural type should be — or matching a differently-ordered type requires a copy, which breaks the zero-cost property the whole data model rests on. Accepted in conversation as a limitation to respect rather than fight.

Writability is what decides whether the copy escape exists. Fields would be writable on the same terms as any struct: by value you get a copy and writes are local; through a (Ptr T) they reach the original. But that settles the layout question rather than sitting beside it. Read-only structural access has an out — the compiler could gather the requested fields into a temporary, and then field order would not matter at all. Writable structural access cannot: it has to alias the real storage, so the layout must genuinely line up. Decide writability first; the layout answer follows from it.

defclass may solve this, for the cases that can afford it. A class has an implementation-defined representation, so the compiler owns the layout and access goes through metadata rather than a fixed offset — field order stops being a user-visible fact. The limit is that a class carries identity and metadata, and a Vector2 should not pay for either. So plain structs still need their own answer and classes only cover the cases where the overhead was already accepted. See item 7.

Also unresolved: the shim generates a C typedef per named struct, and an anonymous struct has no name to generate one from. Either anonymous structs cannot cross the FFI, or the generator learns to name them.

5. A JS backend — for web apps, not for games

The goal, stated in conversation: calling JS libraries, and eventually a hiccup-style DSL for writing web apps in Flan. That is a different product from the wasm target and the two do not compete — games go to wasm, web apps go to JS.

This matters because the obvious objection does not apply. "A JS backend cannot call raylib, so the whole graphics layer would need reimplementing on canvas" is fatal for a game backend and irrelevant here: web apps do not use raylib.

The design fork, and it should be decided first:

• Linear memory — an ArrayBuffer standing in for pointers and structs, with typed-array reads and writes. Fast arithmetic, and every value is opaque bytes, so every call into a JS library marshals both ways. This is asm.js, which wasm exists to replace, and it is precisely wrong for an interop-motivated backend.
• Object mapping — a Flan struct becomes a plain JS object, a Vec becomes a JS array. Slower for tight numeric loops, natural for interop, readable output. For this goal this is clearly the right one, and performance is not the constraint people assume: DOM work is dominated by the browser, not by arithmetic.

The consequence to face early: the memory model does not come along. Object mapping means garbage collection, which means no pointers, no manual free, no arena, no allocator. So this is a dialect, not merely a second backend — some Flan programs will not compile to JS, and that should be named up front rather than discovered. Worth deciding which subset is the JS-targetable one, and whether the checker enforces it per target.

The unexpected upside: the dev loop could work on the web. Conditions, restarts and a break loop all need the ability to run new code in a stopped program. wasm cannot have this — --dev is refused there because reload is dlopen and the browser has no sockets. JS has no such problem: it evaluates new code trivially. A JS target might be the only place the live dev loop and the browser coexist.

Reader conditionals, decided in conversation. A cljc-style subset is accepted, and the author chose Clojure's inline reader conditionals over file-level target naming. Note what that changes: this language has no conditional compilation today, and that absence drove a decision earlier the same day — barf signals on web rather than being compiled out, precisely because "this bit is desktop-only" is not expressible. Note also that the sand-on-web work introduced file-level target replacement for C sources (foo.web.c replaces foo.c), so the repo now has a file-level precedent that this decision deliberately does not follow. Both were put; inline was chosen.

Also wanted, and independent of all of the above: #_ to discard the next form, and repeated (#_#_) to discard that many following forms, as in Clojure. Reader-only, small, and useful immediately rather than whenever a JS backend happens.

One dependency: hiccup is a macro, and the macro expander is blocked on Form being a Flan union, which is union values. The backend can start before that; the DSL cannot.

6. C interop as seamless as Zig's — built, with two decisions left

The mechanism is in (lib/cimport.ml, lib/cjson.ml, vendor/raylib/headers; BUILT.md, "The header is read now"). Settled and not worth reopening: clang's JSON AST dump over a shelled-out clang, never libclang — and Zig has since abandoned linking clang too, for Aro, which strengthens the argument rather than weakening it. The import is bounded by the package's own defstructs rather than by a curated list. Refusals are demotions in Zig's sense: the name exists, cannot be had, and says why at the use site, which Load.refuse_hidden already did for main. Names kebab by a rule that is injective over raylib's 581, and reversibility is by storage — the C symbol is kept verbatim — so the rule never needs an inverse. Where two names do collide, neither takes it.

The evidence: against raylib 5.5, all 16 defstructs and all 172 hand-written declare-c agree exactly. Against the 5.1-dev header also on this machine, ten real differences. Both comparisons stop the build, and a permuted Texture2D or an f64 for a float is caught by name.

What is left is two decisions, and both are the author's.

6a and 6b are answered. See BUILT.md, "The bindings are committed now". It became a code generator whose output is committed, and the 172 hand-written lines were not migrated — for a reason 6b did not reach. The two sections below are kept as the reasoning that got there.

6a. Does reading the header stay a build-time step, or become a code generator?

The cost, measured, with the wrappers pruned by Reach as they already were:

	today	+ 256 imported
release build, warm	0.078s	0.082s
redefinition	31.0ms	46.5ms
dev build, cold	0.649s	0.982s

Release is nearly free and that question is answered. The 15.5ms on redefinition is not nothing on the branch where the dev loop is the priority — it is a 50% increase on the number that lane exists to keep small, and it buys a check of signatures that have not changed since the last build.

So the third option from the original discussion is now the live one: flan import-c already prints declare-c lines, so generate from the header, commit the result, regenerate when raylib moves costs nothing more to build. Explicit in the source, checked against reality, no header read at build time, and the check becomes a thing you run rather than a thing you pay for. Against it: a committed file goes stale silently, which is the failure the whole lane exists to prevent, and "regenerate when the library moves" is a discipline rather than a mechanism.

A middle reading worth considering: keep the build-time check but run it only when not --dev, on the grounds that a release build is where a wrong signature must not get through and a dev build is where 15.5ms is felt. That is the same shape as Reach not pruning dev builds, for a symmetric reason.

6b. Do the 172 hand-written lines get migrated?

The diff is clean, so nothing blocks it on correctness. What blocks it is that migration needs the header present at every build, which means vendoring raylib.h into the repo or requiring raylib-devel — and BUILT.md records "a build needs libraylib linkable and not raylib-devel installed" as a property that was chosen on purpose. That is why vendor/raylib/headers is opt-in (?${FLAN_RAYLIB_H}) today and the hand-written lines are untouched.

Worth noting what migration would actually lose, since it is small but real: the hand-written names are better than the rule's. IsKeyPressed is key-pressed? by hand and is-key-pressed by rule; CheckCollisionRecs is collision-recs?. And an enum parameter imports as i32, because the header says KeyboardKey and nothing tells the importer the package calls that Key — so (rl/key-down? :space) would become an integer at the call site. A migration is therefore not a deletion; it is a deletion plus a kept list of the lines whose face is deliberately nicer than the header's.

Answered: no, and the blocker is not the one above. Vendoring stopped being the question once the output was committed rather than the header — no header is needed at any build. (Vendoring then happened anyway, and for the other reason: vendor/raylib/raylib-5.5.h is committed, headers names it with no ${...} in front of it, and the signature check runs on every build. What that settled was not "can we build without a header" — the committed bindings had already settled that — but "does the check actually run", and behind an opt-in the answer in several worktrees was no.) The count was also smaller than feared: 136 of the 172 are exactly what the rule produces, and the other 36 are expressible as name overrides in bindings.

What actually decides it is that migration would gut the check. Everything the generator emits agrees with the header by construction, so diffing generated output against its own source proves nothing; the hand-written lines have a different author, so they are the only declarations a header can contradict — and all ten of the 5.1-dev differences came from them. Delete them and the signature half of the check silently becomes a tautology. The enum point above survives intact as a second reason: (rl/key-down? :space) keeps its Key parameter only because that line is hand-written.

7. Watching variables

Raised while designing item 1, and deliberately separated from it.

Item 1 shows globals per frame, chosen by what the code in that frame references. That is the right default and it cannot cover the case where the thing you care about is not named by the frame you are standing in — stopped deep in a helper, still wanting to see the grid.

So: a way to say "always show me this", surviving a resume and the next break.

Questions it opens:

• What can be watched. A global is easy — it has a name and an address that does not move. A local is harder: it belongs to a frame that is gone as soon as you resume, so "watch y" either means a different y every time or means nothing. A watched expression — (at velocity 40 12) — is the most useful and the most expensive, since it has to be compiled and run in the program each time, which is what eval-expr already does.
• Where the list lives. Per project, per session, or in the file. A watch list that vanishes when Emacs restarts is one you stop using; one in the repo is one you accidentally commit.
• When it updates. Only on entering a break is cheap and probably enough. Live-updating while the program runs is a different feature — closer to a HUD than a debugger — and worth not conflating.
• Whether it belongs in the break buffer at all, or in its own window that is useful while the program is running, which is arguably where a game developer wants it.

8. Indentation is wrong inside a binding vector

Reported against the let in sand.flan's settle: the second and later bindings indent one column too far.

Cause, found by reading flan-mode.el. flan-indent-function looks at the head of the enclosing form and, when it is one of Flan's body forms (defn, let, if, while, until, dotimes, match, do, loop, defer), indents one past the open paren. Inside a let's binding vector the enclosing open is the [, and the symbol after it is the first binding's name — never a body form — so the check fails and it falls through to Emacs's generic lisp-indent-function, which treats the vector as a function call and aligns continuation lines under the first argument rather than under the first binding.

What it is built on: Emacs's built-in lisp-indent-function, with that one override. Not clojure-mode, which is where the missing piece lives — clojure-mode special-cases binding vectors and aligns them as pairs.

The fix, agreed in conversation: add a binding-vector case to flan-indent-function. When the enclosing open is [ and the form containing it is a binding form, align continuation lines to the column of the first binding, not to the first argument. Bindings are pairs, so the alignment that reads correctly is name-under-name. Check the other bracket users in the same pass — defn parameter lists and restart-case clause parameters have the same shape and almost certainly the same bug.

clojure-mode should be the reference, not the ancestor and not a dependency. The distinction matters and was muddled once in conversation: the proposal is not define-derived-mode on clojure-mode, and not adding an external package. It is to rewrite flan-mode's indentation and font-lock as our own code, ported from clojure-mode's source rather than from lisp-mode's.

The current mode has the wrong ancestor. It is built on Emacs's lisp-indent-function, and Emacs Lisp has none of the shapes Flan actually uses: no vectors as binding forms, no keywords as map keys, no destructuring, no bracket variety. So every rule has to be added by hand and the binding-vector bug is simply the first one hit. Clojure's rules already cover brackets-mean-binding, pairs-align, maps, keywords, #_, and reader conditionals — the last two now wanted in their own right.

Where Flan diverges it diverges deliberately rather than by discovery: defn carries a return type between the parameters and the body; field access is (.x v); field labels are moving from :x to .x, so {.x 1.0} is a struct literal and not a map with symbol keys. Those get written on purpose.

Deriving at runtime stays rejected — it would add an external dependency to a mode that ships inside this repo and currently needs nothing beyond stock Emacs, and that community is mid-transition to a tree-sitter mode, so the base is moving.

9. Instrumenting with (pause) from Emacs, without editing the buffer

Requested, not scheduled. The Clojure model: C-u before an eval marks the form so it stops when it runs. Wanted here for three targets — the top-level form, the last expression, and the form point is inside, so that with the cursor at (+ 1| 1) the program stops at that (+ ...).

What already exists, so this is mostly assembly:

• (pause) itself, landed today in lib/prelude.ml. It is error under a restart-case with a continue clause — a breakpoint is not a compiler feature here, it is what a condition system already gives you. Taking continue resumes at the call.
• flan-eval-defun takes the top-level form's bounds and sends the text; flan-eval-last-sexp sends the sexp before point. Both hand a string to the daemon, which is the hook — the buffer never has to be modified.
• flan-dev--enclosing-head (flan-dev.el:1009) already walks out to the enclosing form to find its head, for eldoc. The third target's hard part is already written.
• current-prefix-arg is already read in one place (flan-dev.el:1429), so the C-u convention has precedent.

So the work is a source-to-source rewrite in Emacs before sending: find the target form's bounds, send the top-level form with that span replaced by (do (pause) <span>). Nothing on the compiler side changes.

The one real difficulty: source locations. Splicing text shifts every line and column after the insertion, so error overlays, layout, the break loop's frame locations and DWARF would all point slightly wrong for an instrumented form. Options, none free: pad the inserted text to preserve offsets (ugly, fragile across newlines); send the instrumentation as a separate field the daemon splices after parsing, where locations are already attached; or accept the drift and say so. The second is almost certainly right and makes this a small daemon change rather than a pure-Emacs one — which also means it should not be built as a pure-Emacs hack first.

Settled: it sticks, like Clojure's. The mark stays until the form is evaluated again plainly. That is the behaviour the workflow expects — you mark it, run the game, hit it repeatedly, and clear it with an ordinary C-c C-c rather than having to re-mark before every run.

10. A second backend for dev builds — study SBCL and drop LLVM there?

Raised at the end of the session, not answered. The proposal: write our own code generator for debug builds and the interactive workflow, keep LLVM for release. The motivation is that evaluating arbitrary code — full functions, real test code, not just constant arithmetic — costs a compile, a file and a dlopen every time.

What SBCL actually does, since it is the model: its own assembler and code generator, emitting machine code straight into the heap. No files, no external tools, no dynamic loading. Code objects are ordinary heap objects and are collected when unreachable. That is also why its disassemble can annotate richly and why it can show machine state — it owns everything.

The argument against, and it is the same one that killed the interpreter. Two backends must agree about arithmetic, struct layout, overflow, evaluation order and calling convention. When they disagree you get works interactively, breaks when shipped — the worst bug class in a live-programming system, and arguably worse than the interpreter version because the divergence is in code generation rather than in semantics you can read.

Note what plan.org's dev-vs-release table already accepts: indirection cells, a shadow stack, version words, code never freed. Those are additive instrumentation on one code generator, not a second one. This proposal is a different kind of divergence.

And the cost is enormous: an instruction selector and assembler per architecture — x86-64, arm64, wasm32 — which is where SBCL's decades went.

The cheaper route to the same goal, and the thing to weigh it against: ORC, LLVM's in-process JIT. Measured today, a 35ms redefinition is 21ms of llc plus ld — external tools generating and linking machine code. An in-process JIT replaces exactly that, with the same IR and therefore the same semantics, no second implementation, and code that can be unloaded. It gets what the proposal wants without the divergence hazard.

plan.org already says ORC "remains addable later behind the same typed IR without touching the language", and the only column text-IR-plus-clang loses in its comparison table is the JIT one.

So the honest position: the thing that would justify revisiting the LLVM decision is not slower eval in the abstract — it is wanting arbitrary interactive evaluation of real code as a first-class workflow. That is now stated as a want. If it stays a want, ORC is the answer and a hand-written backend is not.

Where the shared objects live, since it was asked: Build.workdir () — a per-build temporary directory — with the object cache separately at Build.cachedir (), stable across builds. Nothing sweeps the workdir, and nothing is ever dlclosed, so a long session accumulates both files and mappings. Sweeping at session end is cheap and unrelated to any of the above.

NEXT SESSION STARTS HERE

The agreed plan: investigate putting the compiler inside the running program.

Today there are two programs. The daemon is the compiler. The game is a separate program it launches. They talk over a socket, so everything the editor shows — locals, globals, the stack, evaluation results — has to be packed up in one program, sent across, and unpacked in the other. A large share of what BUILT.md documents exists only to move data between those two address spaces.

Putting the compiler inside the game removes all of that. There is nothing to send, because the compiler can read the program's memory directly. It is what SBCL does, and it is why interactive work in SBCL is instant.

Read, in this order: item 13 (what is actually being decided, and why the order matters), then item 11 (the shape, the two objections and how the author answered them), then item 12 (the specific questions a spike must answer).

First task, and it is research not building: the spike in item 12 — what it costs to link the OCaml compiler into a dev build. OCaml 5.2's multicore runtime has already removed the objection that would have been fatal. If it embeds acceptably, everything else is reachable.

Do not start with the backend question. Merge first, measure, then choose. Item 13 says why.

---

11. One process, not two — the biggest open question

Raised at the very end of the session and not resolved. It subsumes items 9 and 10, and possibly several decisions recorded elsewhere as settled. Do not treat any of those as closed until this is answered.

The claim: the daemon and the running program should not be two processes. SBCL is one image; Clojure is one image; in both, the interactive workflow is cheap because nothing is transported. The author's reading is that the split here was forced by not having threads — and that is worth checking, because threads are already in use: the agent runs a listener thread and a loader thread inside the program today.

The concrete shape, if it were done: load the program into the daemon's process as a shared library rather than launching it as a child. The reload primitive already dlopens modules; the program itself would be one.

What would evaporate, and it is a lot. No socket and no wire protocol. Cell updates become plain pointer stores. Locals and globals become memory the compiler can read directly — no render thunks compiled per inspection, no 4K result cap, no seqlock, no snapshot copying, no generation stamping. The macro expander stops needing the compiler's own dlopen. flan reload stops being a fresh process paying startup every time. A large amount of machinery recorded in BUILT.md exists only to move data between two address spaces.

What is given up, in order of sharpness:

1. Crash isolation. A segfault in the game currently cannot touch the compiler. plan.org's jank note — corrected earlier for citing the wrong mechanism — concludes "we are safe from the repro because we compile out of process". In one image a bad pointer takes the session with it. Lisp users live with this and mostly accept it; it is a real loss and should be chosen knowingly, not discovered.
2. macOS requires a window on the main thread. "Run raylib on a separate thread" is fine on Linux and not portable. The inversion — compiler on the side thread, window on main — probably works, but it needs deciding rather than assuming, and it changes who owns the main loop.
3. The OCaml runtime and the game in one address space: signals, SIGSEGV handling for the break loop, the OCaml GC running while game code holds raw pointers into its own arenas. Each is probably fine; none is free.

What it does NOT fix, and this is where an earlier answer in this session was wrong. A hand-written debug backend was discussed as the way to make evaluation instant. The correction: a new backend removes llc and ld — about 21ms of a 35ms redefinition — but not the transport, because the transport exists due to the process split, not the backend. Conversely, merging the processes removes the transport but not the code generation. They are independent, and merging is the larger prize of the two.

Decisions to revisit if this is taken: the watch design (push was chosen partly because polling costs a compile); the 4K result cap and its seqlock; the snapshot machinery and its generation stamping; the render-thunk-per-inspection design for locals and globals; the inspector's inability to retain a value; and whether ORC or a hand-written lowering is needed at all.

Both objections were answered by the author, and the second reframed the design — record the corrected shape.

Crash isolation: accepted knowingly. Conditions already catch what the language signals — a failed bounds check, an exhausted allocator, a missing file — and those keep stopping in the break loop in one process exactly as they do now. What they cannot catch is a real memory fault: a bad pointer through the FFI, a use-after-free in an arena. That is an OS signal, not a condition, and it takes the image with it. Same deal every Lisp makes.

macOS: invert the relationship, and this is the right shape rather than a workaround. Not "load the program into the daemon" as sketched above — the opposite. The agent is already a server running inside the program. So the compiler moves into the program's process, beside the listener that is already there. The game starts, takes the main thread for its window, and the compiler and listener run on a side thread. That is SLIME's model: you start the image, it serves, the editor connects.

This also fits a distinction plan.org already draws — dev and release builds are deliberately different. A dev build links the compiler in; a release build ships neither compiler nor server. The same split SBCL has between a development image and a delivered executable.

One thing that gets better rather than merely cheaper: a break loop in the same process as the compiler can offer restarts that recompile and retry, with no transport in between.

Suggested next step: the macOS question is answered, so start instead with what it costs to link the OCaml compiler into a dev build — the OCaml runtime and the game sharing an address space, signal handling for the break loop, and the GC running while game code holds raw pointers into its own arenas. It is the only one of the three costs that could make the whole idea non-portable, and it is answerable by reading raylib and GLFW rather than by building anything.

12. What it costs to put the OCaml compiler in the game's process — and whether to port the compiler

Follows from item 11. Investigate before deciding anything; this entry is the brief, not the answer.

One fact established already: this project is on OCaml 5.2. That is multicore OCaml — real threads, domains, no global runtime lock. It removes the objection that would have been fatal: the runtime can sit on its own domain beside the game's threads rather than serialising everything.

What to find out:

• Linking OCaml into a C/native binary. ocamlopt -output-obj and friends are supported; how well, with dune, with the C stubs this project already has (dynload_stubs.c, the runtime, the shims)?
• Signals. The OCaml runtime installs handlers. The break loop wants SIGSEGV for the crash case, and the agent already handles socket work. Who wins, and can they coexist?
• The GC and raw pointers. OCaml's collector moves its own heap; Flan's arenas and Vecs are raw memory OCaml never sees. That should be fine — they do not point at each other — but "should be" is not an answer.
• Startup cost and binary size for a dev build that links the whole compiler.
• Whether the main thread can be the game's while the compiler runs on a domain (see item 11's macOS note).

And the larger question the author raised: port the compiler to another language?

The honest end state for a Lisp that wants one image is self-hosting — the compiler written in Flan. That is what SBCL is, and it dissolves this entire question: no foreign runtime to embed, no GC to reconcile, no signal conflict, and the compiler becomes the largest test of the language. It is also an enormous undertaking and needs macros, unions, Map and a string library first — three of which landed today.

Intermediate options if OCaml proves genuinely hostile: C or C++ (no runtime to embed, painful to write a compiler in), Rust (no GC, embeds cleanly, large rewrite), or Zig (same, and its own C-interop story is already being read for item 6). None of these should be entertained on aesthetics. The bar is a specific, demonstrated obstacle to embedding OCaml 5.2 — and the multicore runtime means the most likely obstacle has already gone away.

Order of work: answer the embedding questions above with a spike, not a rewrite. If OCaml embeds acceptably, item 11 is buildable and no port is needed. Self-hosting stays a long-term ambition rather than a prerequisite.

13. The two questions are independent, and the order matters

The end of the session, and the clearest statement of what is actually being decided. Read this before items 10, 11 and 12, which were each written mid-argument.

Two separate decisions, repeatedly conflated in conversation:

1. One process or two — whether the compiler runs inside the program's process. This is about transport: how compiled code and inspection data move between the compiler and the running program.
2. How machine code is produced and installed — the current llc + ld + dlopen, libLLVM's in-process JIT (ORC), or a hand-written lowering. This is about code generation.

Neither implies the other. A hand-written backend removes llc and ld (about 21ms of a 35ms redefinition) and leaves the transport. Merging the processes removes the transport and leaves code generation exactly as it is. Merging is the larger prize.

And there is a third option for question 2 that keeps getting lost: change nothing. In one process you can still write a .so and dlopen it. You lose nothing that works today, and you still delete the socket, the wire protocol, the 4K result cap, the seqlock, the snapshot copying, the render-thunk-per-inspection, and flan reload paying process startup every time.

So the order is:

1. Spike the embedding (item 12). OCaml 5.2's multicore runtime removes the objection that would have been fatal. If it embeds acceptably, everything below is reachable.

2. Merge the processes, keeping the existing code path. Biggest win, no new dependency, deletes the most machinery.

3. Measure what is left. With transport gone, a redefinition is whatever code generation costs. That may simply be fine.

4. Then decide the backend, on evidence. libLLVM's JIT means linking the library plan.org rejected — for reasons that were right at the time, when the JIT was worth ~13ms and nothing else. In one process it is worth the file, the load, the unbounded accumulation and cheap evaluation of arbitrary real code, which the author has since named as a first-class want. The cost is also more bounded than "breaks routinely" suggests: the bindings track LLVM major versions, so it is a pin you own and upgrade deliberately.

   A hand-written lowering buys the same speed plus no dependency and a disassembler you own, at the cost of an instruction selector per architecture, maintained forever. The drift risk is real but testable — the @sanitize alias already builds 28 programs twice and compares output and exit status, which is exactly the harness two backends would need.

The principle to keep: the original decision was made on a measurement and it held up for months. Make this one the same way — merge first, measure, then choose. Do not pre-empt step 4 at step 1.

Also settled in passing: a release build can include the compiler and server and simply not be connected to, the way an SBCL executable can. The dev/release split is a build flag, not an architectural constraint.

And one correction to keep: OCaml's own wasm support is irrelevant to any of this. OCaml is the compiler's implementation language; Flan programs reach wasm through LLVM. The two only meet if the compiler should run in a browser, which is not a goal.

14. The embedding spike, answered: OCaml 5.2 goes into a Flan dev build, and nothing objects

Item 12's brief, run. Feasible. No obstacle was found that argues for porting the compiler, and the one expected to be sharpest — signals — turned out not to exist on the platform measured.

The apparatus is spike/embed/: four shell scripts and sixteen small sources, deliberately not a dune target, driving ocamlfind and clang by hand against the flan.cmxa dune already builds. bash spike/embed/run.sh reproduces everything below; sig.sh, symbols.sh and merged.sh each answer a question on their own. Nothing under spike/ is wired into the build, and dune test --root . is green either side of it.

The headline: one binary, and it compiles itself

spike/embed/merged.sh builds a single executable out of, in one clang link:

• the emitted Flan program (test/programs/edn.flan, --dev, @main renamed to flan_program_main),
• runtime/flan_rt.c, runtime/flan_dev.c, vendor/agent/flan_agent.c,
• and the entire OCaml compiler as one -output-complete-obj object.

It runs. A C main() holds the main thread and runs the Flan program there; caml_startup happens on a pthread beside it, next to where flan_agent_start already puts its listener. The compiler inside the binary then compiles edn.flan — the source the program itself was built from — and emits 336,579 bytes of LLVM IR. That single result answers questions 1 and 2 together; the ladder of smaller probes underneath it is support, not evidence in its own right.

Nothing is wired up. The two halves share an address space and do not speak to each other. That is the point: the question was whether they can, not what they would say.

Question 1 — does OCaml link into a native binary here?

Yes, and with less friction than expected.

• ocamlopt -output-complete-obj is the one to use, not -output-obj: it bundles the runtime, so there is no hunt for libasmrun. The final link needs -lm -lpthread -ldl and, on 5.2, -lzstd — 5.x's marshaller is compressed, and the missing ZSTD_* symbols are the first thing a naive link fails on. That is the whole of the surprise.
• dune is not in the way, because it does not have to be involved. dune builds lib/flan.cmxa as it does today; the -output-complete-obj step consumes that artifact afterwards. No dune rule had to change, and lib/dune and bin/dune are untouched. A real merge would want a dune rule to drive that step, but the spike shows the artifact boundary is clean, which is the part that could have failed.
• The existing C stubs come through. lib/dynload_stubs.c was taken verbatim from the unmerged 9e0ae3a dlopen branch and compiled into the same object; from inside the embedded runtime, flan_mem_alloc/poke/peek round-trip correctly and dlopen+dlsym work. -output-complete-obj carries a foreign_stubs-shaped C file through without special handling.
• No symbol collides (symbols.sh). Flan's own C — flan_rt.c, flan_dev.c, flan_agent.c, dynload_stubs.c — defines 211 symbols; libasmrun.a defines 1,379; the intersection is empty, and the four Flan files do not collide with each other either. Worth checking rather than assuming: those four are compiled into two different processes today, and merging puts them in one link for the first time.

Question 2 — threads, and who owns the main loop

OCaml 5.2 is confirmed multicore, and Domain.recommended_domain_count () reports 16 here. A spawned domain does real work in parallel with the main thread. The objection that would have been fatal is gone.

The macOS shape works, and it was tested as the thing that matters rather than as "can OCaml use threads":

• caml_startup can be called from a C-created, non-main pthread, while main() goes on to a loop it does not leave. harness4.c runs a 298-frame mock game loop on the main thread that never once enters OCaml.
• A second C thread — one the runtime never created, which is exactly the agent's listener — can call into OCaml after caml_c_thread_register(), bracketed by caml_acquire_runtime_system/caml_release_runtime_system. It compiled the same program successfully from that thread. This is the specific capability the merged design needs from flan_agent.c, and it exists.

The spike could not test macOS. Nothing here is macOS-specific — the inversion is a portable pthread arrangement — but it is a Linux measurement.

Question 3 — signals

On Linux/amd64, OCaml 5.2 installs no signal handlers at all. Not SIGSEGV, not SIGINT, not SIGFPE, not SIGPIPE, nothing. The expected conflict does not exist.

sig.sh sweeps fifteen signals from inside the runtime at four moments — before caml_startup, at module init, from inside a spawned domain, and after Domain.join — and every one is SIG_DFL. The reading has to be taken from inside OCaml rather than from C after caml_startup returns, because OCaml 5 starts domains later; the first pass got this wrong and read SIG_DFL for the wrong reason. A plain ocamlopt executable was built as a control and behaves identically, so embedding changes nothing about signals.

The reason is structural, not incidental: OCaml 5 detects stack overflow with an explicit stack-limit check and calls caml_raise_stack_overflow directly, rather than with a guard page and a SIGSEGV handler. nm on libasmrun.a shows sigaltstack referenced but no SIGSEGV handler defined, which matches.

So the break loop can take SIGSEGV outright, and it does not have to install first or last — order does not matter when there is nothing to displace. Measured directly: with the break loop's handler installed, deep OCaml recursion still raises Stack_overflow normally, and a genuine fault at 0x10 reaches the break loop's handler with si_addr correct. Chaining was implemented and compared; it is unnecessary here, but harness5b.c keeps it, because it is what the merged build should do on any platform where the sweep comes back non-empty.

The caveat, and it is the one thing in this report to re-check rather than trust: this is x86_64-pc-linux-gnu only. macOS/arm64 OCaml 5 was not testable here, and item 11 raises signals in the same breath as macOS. Re-run sig.sh there before relying on it. flan_agent.c needs nothing from this either way — it sends with MSG_NOSIGNAL throughout rather than depending on a SIGPIPE disposition.

Question 4 — the GC and raw memory

Confirmed, and the confirmation is narrow on purpose. An 8 MiB arena was filled with a checkable pattern and 64 raw interior pointers taken into it; OCaml then allocated 26.4 million words, took 7 major collections and a full Gc.compact (). Afterwards: the arena base is unmoved, 0 of 8,388,608 bytes altered, 0 of 64 interior pointers invalidated.

What that proves is that the collector traces its own roots and foreign memory is invisible to it. Arenas, Vecs and Maps are safe because OCaml never learns they exist.

What would break the assumption, stated so it is not rediscovered the hard way:

1. Storing an OCaml value in Flan memory — an arena, a Vec, a global — across any allocation. The collector will move the block and will not update that word, because it is not a root. caml_register_global_root (or the generational one) is the only way that is legal, and it is a rule the merged design has to hold: the boundary passes pointers and scalars, never values into Flan storage — the same rule dynload_stubs.c already states for its own reason.
2. A value held in a C local across a call that allocates, without CAMLparam/CAMLlocal. Ordinary OCaml-FFI discipline; it applies to every stub the merge adds.
3. Long C work on the compiler thread without releasing the runtime system — which corrupts nothing but stalls whichever domains want a stop-the-world. llc and ld are execs and would want caml_release_runtime_system around them.

Question 5 — what it costs

	bytes
edn.flan, release build	71,824
edn.flan, dev build, as built today	114,296
the same dev build with the whole compiler linked in	4,257,624
what the compiler adds	~4.14 MB

Startup: caml_startup takes 0.58–0.72 ms across six runs — the runtime coming up and every module initialiser in the compiler running. Measured with clock_gettime around the call itself, not time(1) on the process, because exec and dynamic linking are paid today anyway.

The size reads as 37x, and that framing is misleading. A dev session today runs two binaries, and the daemon alone is 4,815,368 bytes. The merged dev build is smaller than today's compiler process by itself, and there is one of it instead of two. Sub-millisecond startup and ~4 MB is not a cost worth designing around.

One more number, recorded because item 13's step 3 will want it and for no other reason: a full in-process compile of edn.flan — read, parse, load, typecheck, emit — is 12.1–12.5 ms, warm and cold alike, with llc and ld excluded because they are separate processes. That is where the remaining cost sits once transport is gone. It argues for nothing; item 13 says the backend is decided at step 4 on a measurement taken at step 3, and this is not that measurement.

What this does not answer

• The agent's handlers are a port, not a recompile. harness4.c proves the pattern — a C-created listener thread can register with the runtime and call OCaml. It does not port flan_agent.c's handlers, which today answer requests out of the program's own memory and would instead be calling into the compiler.
• Crash isolation is gone by construction. Not a finding; item 11 already accepts it knowingly. A bad pointer through the FFI takes the session, and conditions still catch everything the language signals. Noted only so this entry stands alone.
• macOS, for signals and for the main-thread inversion. See above.
• The backend. Untouched deliberately.

The order the real work goes in

1. Make the -output-complete-obj step a dune rule, producing the compiler-as-object that a dev build links. This is the only build-system work, and lib/dune/bin/dune did not need changing to prove it.
2. Land the dynload_stubs.c branch (9e0ae3a, currently reverted). The merged build needs the same pointer-and-scalar boundary, and it is already written.
3. Invert the startup: the Flan program keeps main(), and flan_agent_start also brings up the OCaml runtime on its side thread. merged_main.c is the sketch.
4. Port the agent's handlers from answering out of the program's memory to calling the compiler directly — this is where the socket, the wire protocol, the 4K result cap, the seqlock, the snapshot copying, the generation stamping and the render-thunk-per-inspection all get deleted. It is the bulk of the work and the whole of the prize.
5. Keep llc + ld + dlopen exactly as they are. Item 13's third option. Nothing here argues against it.
6. Measure what is left. Then, and only then, item 13 step 4.

15. The backend spike, answered: feasible, and the obstacle is not the one anyone expected

Item 13's step 4, run early and deliberately out of order, as a spike rather than as a decision. Feasible. A function written in Flan goes through the ordinary frontend, is lowered to x86-64 by hand, is written into an mmap and is called, and it answers correctly. That took an afternoon and no reference material beyond objdump.

The apparatus is spike/backend/: x86.ml (an instruction selector), jit_stubs.c (three calls OCaml cannot make for itself, plus the C side of the ABI probes), probe.flan, driver.ml and run.sh. bash spike/backend/run.sh reproduces everything below. There is no dune file under spike/, nothing is wired into the build, and dune test --root . -j 1 is green either side of it.

The headline is not the arithmetic. Ten checks pass, and one of them failed first and mattered: a C callee that does a 16-byte aligned spill and reports whether it was entered aligned. Called plainly it passed; called from inside a binary operator it returned -1. The evaluator spills its left operand across the right operand's evaluation, so a call written in the right operand runs with rsp 8 bytes out. That is the raylib crash, reproduced on day one of a backend that does almost nothing. It is fixed with a depth counter and an assertion, and it is the clearest single argument that this work is tractable but unforgiving: nothing about the wrong version looked wrong, and only a probe built to catch it caught it.

Question 1 — does one function work end to end

Yes. The frontend is the real one — Reader, Parse, Load, Check — so what is lowered is the same Tast.fn emit.ml gets, not a literal typed to make the exercise come out. Eleven of the 83 functions in a trivial program (the prelude is most of them) lower with no special handling, including space?, digit? and upper-ascii, which nobody wrote for this.

The emitter is the trivial one the brief allows: every slot is a stack slot at rbp - 8(i+1), every value is computed into rax, a binary operator spills its left operand. Two registers, no allocator, no liveness. spike-add is 58 bytes where clang -O0 would spend about 20, and that is the correct trade for a debug build.

Proved by comparing numbers, not by reading bytes. SPIKE_DISASM=1 disassembles the buffers that ran, and that is a debugging aid kept out of the pass/fail path on purpose: a disassembly that reads correctly beside a function answering 656 when it should answer 650 is the normal outcome of hand-encoding.

Question 2 — the real shape of the work

Layout is already owned, and this is the best news in the report. emit.ml's lay / lay_fields / payload_lay compute C struct layout — offsets, padding, tail padding, the union payload blob — because DWARF needs member offsets as integer literals and getelementptr cannot supply one. They are acceptance-tested against LLVM's own answer for the same struct type. So the drift risk item 10 fears most, two backends disagreeing silently about where a field is, does not arise: there is one layout calculator and a new backend calls it.

Against tast.ml's expr_kind, in four buckets:

	nodes
Done in the spike	Int Bool Local Do Let If Return Set/Plocal, arithmetic, comparison, bitwise, Call, Rt
Mechanical	While Break Continue (the jump patching exists), Global Str Zero Uninit, the rest of place, Field Deref Addr, Arr, Some_ None_ UnwrapSome, Match on a tag
Bulky, not hard	floats — a second register file, SSE encodings, Cast's eight conversions, and the SSE half of the calling convention. Perhaps a third of the total instruction work for a small fraction of the programs
Fiddly	aggregate copy on assignment (a struct store is the copy spec-memory.md requires), Make MakeCase CaseField over the payload blob, CallPtr, FnAddr's three cases and the cell load behind Fnval
No plan	Handled Signal RestartCase InvokeRestart WithAlloc, the transfer-channel guard after every call, the landing pad, and fdefers on the transfer exit path

The last row is the one to take seriously. The spike never emitted a guard or a pad, and the guard is on every call site in the real thing — emit.ml's guard, current_pad, emit_restart_case and emit_with_alloc are several hundred lines of control flow that a second backend reimplements from the spec rather than copies. Conditions are not an advanced feature to defer: spec-conditions.md is load-bearing in the prelude already.

Question 3 — the SysV boundary, and the obstacle nobody named

The C boundary is the easy half, and BUILT.md is why. check.ml rejects an aggregate in a declare signature and the generated shim flattens every struct, so no Flan-emitted call ever passes one to C. A string or slice crosses as ptr+len — two arguments, which is the only counting subtlety. The spike calls an eight-argument C function correctly, including the two that go on the stack, and sets al for the variadic case. No aggregate classifier is needed for raylib. That is a large piece of plan.org's "three classifiers to write and keep correct forever" that simply does not apply.

The hard half is Flan calling Flan, and it was found by reading signature. That function spells each parameter with ll ty and flattens nothing. Emitting a trivial program and looking at the define lines:

define i64            @"flan.take-slice"(%slice %p0, i64 %p1, ptr %xfer)
define { i8, i32 }    @"flan.index-of-i32"(%slice %p0, i32 %p1, ptr %xfer)
define { i8, float }  @"flan.min-f32"(%slice %p0, ptr %xfer)
define { i8, %slice } @"flan.split-next!"(ptr %p0, ptr %xfer)
define %vec           @"flan.filter-i32"(%slice %p0, ptr %p1, ptr %xfer)

The prelude is wall to wall aggregates by value. And what LLVM does with them, measured by objdump on clang's own output rather than read off a table:

• %slice argument → rdi:rsi, two registers, and the next argument shifts along.
• { i8, i64 } return → tag in al, value in rdx.
• { i8, float } return → tag in al, value in xmm0.
• %vec return (six words) → a hidden sret pointer in rdi, the real arguments shifted along behind it, and the pointer returned in rax. That pointer does not appear in the define line at all.

The last two are not the C ABI, and that was checked against a control rather than asserted — the same three shapes written in C, compiled by the same clang at -O2, beside the same shapes written as first-class IR aggregates:

shape	from C	from IR
{ i8, i64 }	al + rdx	al + rdx — agree
{ i8, float }	packed into rax (movd/shl/or)	al + xmm0
{ i64, i64, i64 }	sret pointer in rdi	rax + rdx + rcx

The 24-byte case is the striking one: C spills to memory through a hidden pointer, and the IR form returns it in three registers, one of which — rcx — the SysV ABI never uses for a return value at all. Somewhere past that, LLVM does switch to sret, which is what %vec gets.

So the internal calling convention is not the C ABI and is not specified anywhere. It is whatever LLVM's backend does with a first-class struct, discoverable only by disassembling. That is the sharpest obstacle in this report, and it is sharper than raylib for three reasons: the reference is an implementation rather than a document; the failure mode is a garbage field rather than a link error; and it is not stable by contract across LLVM versions, which is exactly the coupling plan.org chose text IR to avoid.

It also forces the decision that determines everything else:

1. Redefinitions only — the custom backend emits a new body into an LLVM-built host. Incremental, testable one function at a time, and the path that fits the dev loop. It requires matching LLVM's aggregate convention bit-exactly, including the mixed integer/SSE case above.
2. The whole dev build — the custom backend owns both sides and picks the convention: every aggregate by pointer, nothing classified, done. No matching problem at all. But it needs complete node coverage on day one, conditions included, and there is no partial version that runs.

The spike leaned on option 1 without noticing, because the probe called C and C is the flattened half. A real attempt has to choose deliberately.

Question 4 — where the language leans on LLVM instead of defining itself

The audit, which stands on its own whatever happens to the backend. Each row is drift you do not get if the language answers it.

	today	defined?
Integer overflow	no nsw/nuw, "arithmetic wraps (plan.org, Types)"	yes
Shift count	masked to the operand width; a literal out of range is rejected by check	yes
Evaluation order	map_lr, and the comment says left-to-right is required, not a preference	yes
Division by zero	nothing. prelude.ml calls a remainder by zero "immediate undefined behaviour" and routes around it	no
INT64_MIN / -1	nothing, and it is a separate case. LLVM says undefined; x86 idiv raises SIGFPE	no
f64 → i64 out of range	fptosi, undefined in LLVM; x86 cvttsd2si answers the "integer indefinite" value	no
Uninit	emitted as poison	no, and see below
unreachable	after a noreturn call, and after an exhaustive match	no
Alignment	no explicit align on loads and stores; LLVM uses the type's ABI alignment	implicitly, via lay

Two are worth more than a table row.

Uninit → poison is the one that actually bites, and it bites in the direction item 10 fears. A hand backend gives a stable garbage value: whatever the stack slot held. LLVM's optimiser may reason from poison and delete the code that reads it. So (uninit) is the one construct where the two backends are supposed to differ and where "works in dev, breaks when shipped" is the expected outcome rather than a bug. The language should say what reading an uninitialised value means before a second backend exists, not after.

unreachable is the cheap one. The spike emits ud2: a defined SIGILL at the instruction that fell through. LLVM's unreachable is undefined behaviour and licenses the optimiser to delete the path. Defining it as a trap costs two bytes and turns a class of miscompile into a crash with an address.

None of this needs a backend. It is a session with plan.org and six check.ml cases.

Question 5 — unloading code, which is the prize

The shadow stack is better than expected and still not sufficient, and the two halves of that are separate questions.

The running half — and the shadow stack does answer it. The frame push in emit_fn is inside a plain if m.dev and is not gated: only the slot table is gated on named && n > 0, and a function with no named slot still pushes a frame with a null slots. So every active Flan function in a dev build is on the chain, lifted handler clauses included, and each frame points at the flan_fninfo belonging to the module it was compiled into — so the pointer identifies not just the function but which body. "No frame on the chain names this body" is answerable today, with no DWARF and no unwinder.

Two caveats on that half. The pop happens before leave; ret, so a body is briefly executing with no record — irrelevant if reclamation happens at a safe point on the same thread, fatal if another thread reclaims while the game thread is returning. And the chain is a plain global, not thread-local, which flan_dev.c states and justifies.

The pointed-into half, which the shadow stack cannot see and which is the actual reason nothing is dlclosed today. BUILT.md is explicit, and it is not the reason the brief assumed: "a cell holds an address inside a module's text; unloading it leaves every call site pointing at unmapped memory. The rule is about being pointed into." Owning the code object answers most of this — you own the cells, so redefinition drops the old body's last cell reference — but not all of it:

• FnAddr (Fnval n) loads the cell and yields a raw body address, which can then be stored in a struct, a Vec or a global. Nothing records that it happened.
• FnAddr (Flanfn _) and (Rtfn _) bypass the cell by design — tast.ml says they "must never take that path" — so a Map's hash and equality pair and a handler-bind clause's address are raw pointers into a specific body, held in data.

So: frames are tracked, escaped code pointers are not. Unloading needs a rule the language does not have yet. The cheapest honest one is deferred reclamation — retire a body when no frame names it and an epoch has passed with no new capture — and the cleanest is to make a function value a cell pointer rather than a body pointer, which costs one indirection on CallPtr in dev builds and makes the whole question go away. That second option is worth writing down now whatever happens to the backend, because it is a change to what a Fn value is.

The verdict

Feasible, unforgiving, and not the next thing to do.

Feasible: the instruction selection is easy, layout is already owned and tested, the C boundary is already flattened, and one function ran on day one. Nothing here argues the way item 10 feared — the divergence hazard is real but it is concentrated in three named places (Uninit, division, the float cast), not spread through the whole of arithmetic.

Unforgiving: the internal aggregate convention is defined by LLVM's implementation and not by any document, the alignment rule is invisible until raylib crashes somewhere else, and conditions are a second full implementation of spec-conditions.md rather than a port.

Not next: item 13's order still holds, and the spike does not disturb it. Transport is 41µs of a 21ms redefinition and code generation is 19 of the 21, so the speed case remains what item 13 said it was. What this spike adds is that the introspection case is also not free — unloading needs a rule about escaped function values that nothing in the language has, and that rule is worth having whether or not a backend is ever written.

What to do first if it went ahead, in order, and the first two are worth doing regardless:

1. Define the six undefined cases (question 4). No backend required, and every one is drift avoided rather than drift managed.
2. Decide what a Fn value is — body pointer or cell pointer — and write it down. This is the unloading question and it is a language question, not a backend one.
3. Choose option 1 or option 2 from question 3, deliberately. Everything else follows from it.
4. Only then, and only if 3 says so, grow spike/backend/x86.ml from the node table in question 2 — floats first, because they gate most of the prelude, and conditions last, because they are the only row with no plan.

16. The dev backend, wired: a whole program compiles through x86.ml and runs, and conditions were never in the way

Item 15's step 4, taken. lib/x86.ml was 502 lines of encoder and frame model with nothing calling it. It now lowers a whole Tast.program to an assembly file, and flan build --x86 hands that file to the same clang invocation the LLVM path uses, against the same runtime objects, the same generated shim and the same linker arguments. The flag is off by default and refused in combination with --dev, --debug, --sanitize and every wasm target. LLVM stays the release backend and the default one; nothing on the existing path changed. dune test --root . is green either side.

41 programs out of test/programs build through it, and 40 of them print exactly what the LLVM build prints. The 41st is bounds.flan, and it diverges on purpose — see question 4.

Question 1 — which program, and what did it actually cost

Measured with spike/backend/hist.ml before anything was written, over candidates, not guessed. The one picked is spike/x86/p3-fizz.flan — a dotimes, a call, an if, a remainder, two string literals and print/println:

p3-fizz.flan: 2 reachable fns
  Call 1   Do 4    If 1    Int 15   Let 2    Local 7   Prim 14
  Set 1    Str 3   While 1          place/Plocal 1
  prim/Add 2   prim/Bytes 3   prim/Cast 1   prim/Eq 1
  prim/I64ToBytes 1   prim/Lt 1   prim/Rem 1   prim/WriteStdout 4

Nineteen rows, two reachable functions after Reach prunes, and no Signal, no Handled, no RestartCase, no Make, no Field, no allocator. The same is true of a program that only loops and prints: zero of each.

That is worth stating flatly because the expectation going in was the opposite — that the smallest useful program carries one of each condition node, and that printing a single value drags in Str, Make, Field and Call because the prelude builds a slice to do it. It does not. Printing an integer is Cast → I64ToBytes → WriteStdout, three prims and no call at all; flan_i64_to_bytes renders into a static buffer in flan_rt.c and hands back a slice. Printing a string literal is Str → Bytes → WriteStdout, and Bytes is a non-instruction because a string and a [u8] are the same two words.

What does drag conditions in is the bounds check and the allocator, and neither is a Tast node. check_at and check_slice call flan_bounds_error with the transfer channel and then guard; Rt flan_vec_at takes the channel too. So it is m.checks and the container runtime that make a program need the condition machinery, not looping and not printing — and because both are emit-time constructs rather than IR nodes, hist.ml cannot see either. That is the correction to item 15's node table: the "no plan" row is not reached by writing a loop, it is reached by writing (at a i).

Question 2 — what runs

program	what it is for
spike/x86/p1-exit.flan	main returns 0. The assembly path, the runtime link, a real executable.
spike/x86/p2-loop-print.flan	a dotimes that prints — the smallest program that does something
spike/x86/p3-fizz.flan	the measured target: a call, an if, %, two string literals
spike/x86/p4-convention.flan	the internal calling convention, which p3 does not touch at all
spike/x86/p5-core.flan	a global with an initialiser, recursion, break, continue, the bitwise family, unsigned shifts, both directions of every conversion

p4 is the one that matters most, because item 15 named the internal aggregate convention as the sharpest obstacle in the whole report. It passes a struct by value, returns a struct by value, puts an f32 through the SSE half, calls an eight-argument function so that two arguments go on the stack, and passes a slice — and it agrees with LLVM. The obstacle really did dissolve the way the header claims: a dev build is compiled entirely here and a release build entirely by LLVM, the two never meet in one process, so the convention is ours to pick. Every aggregate goes by pointer, an aggregate return is a hidden pointer in the first integer register returned in rax, and there is no classifier in the file. Nothing had to be discovered by disassembling clang.

Over test/programs (111 files):

	n
built through --x86 and matched the LLVM build's output and exit status	40
built and diverged — bounds.flan, by design	1
refused by name: a node this backend does not lower	40
no main (package and library fixtures)	6
do not compile at all (the checker-error fixtures)	22
never terminate on their own (dev-loop, dev-watch)	2

So of the 81 programs that compile, have a main and finish, 41 went through the hand-written backend and 40 were byte-identical in output. That includes edn.flan — sixty lines of output from a hand-written EDN reader with unions, options, nested collections and a fixed-depth balance stack.

Proved by comparing output, never by reading bytes. The script builds both ways and diffs stdout and the exit status; objdump was used only after a program already had the wrong answer. Item 15 is right that this is the only honest order.

Question 3 — the two bugs, and both are the shape item 15 predicted

A (set (.x (at pts 0)) 1.5) wrote into a copy. lvalue had no case for At, so it fell through to "evaluate it", and the store landed in a temporary while the array kept its zeros. emit.ml has this as addr's own At case. One line. array-ctor.flan found it, and it found it as a segfault several statements later.

A discarded value was stored over the return address. This is the better one. A form whose value is thrown away was handed a sink, and the sink was spelled as an address — rbp+0. That is the saved rbp, and rbp+8 is the return address, so a non-void form written in statement position stored straight over both; a 16-byte slice did it in one rep movsb. edn.flan crashed by jumping into .rodata, several statements after the mistake and in a different function, and the assembly at the jump read perfectly. The sink is now compared by identity and never used as an address: anything with a value that is handed it gets a frame temporary instead.

The second bug cannot exist on the LLVM path, and that is the general point. LLVM has no notion of "store this value nowhere" — an unused SSA value is simply unused. Every construct this backend has that LLVM does not is a place where a bug can live that the LLVM backend's own testing can never have covered.

Against that, the thing item 15 was most worried about did not happen: nothing went wrong with the frame or the stack alignment. rsp is written exactly twice — one rounded sub in the prologue that covers the temporaries and the outgoing-argument area together, and leave — so rsp % 16 == 0 at every call site is a property of one subtraction rather than an invariant every case maintains. The spike's worst bug has no door to come in by, and p4 calls an eight-argument function to prove it.

Question 4 — where the two backends now differ, and it is three named places

Item 15's question 4 listed nine undefined cases. Three of them are now real divergences with a build on each side, and they should be written down before anyone uses this for anything.

	LLVM	here
A bounds violation	flan_bounds_error signals; a restart-case can catch it; bounds.flan exits 134	no check at all; bounds.flan exits 139
(uninit)	poison, and the optimiser may reason from it	whatever the stack slot held — stable garbage
an exhausted match, a noreturn call	unreachable, undefined	ud2 — a defined SIGILL at the instruction that fell through

The first is the one that matters, and it is not a footnote: --x86 is silently a --no-bounds-checks build. It is silent because it is not a decision the backend made — check_at signals, signalling needs the channel and the guard, and there is no guard here, so there is no check. bounds.flan is the direct evidence and edn.flan's 33-brackets- against-a-32-deep-stack case is the second. Anyone reaching for this flag on a program that indexes anything should know that the trap is gone.

The other two are improvements and cost nothing. ud2 in particular is two bytes and turns a class of miscompile into a crash with an address.

Question 5 — conditions, which is still the row with no plan

They were not reached, and converting that into a checked precondition is the most useful thing in this report.

There is no transfer guard after a call here, no landing pad and no transfer exit. emit.ml emits a guard after every call; this emits none. What makes that sound is a whole-program argument rather than a hope: if nothing in the reachable set can ever write the channel, no call can ever return with it set. So check_no_transfer walks the linked program once per build and stops it — with the node's name and the function it is in — the moment it finds a signal, an invoke-restart, a restart-case, a handler-bind, a with-allocator, or the two Rt symbols whose bounds check signals.

That is what the 40 refusals are:

27  restart-case        4  handler-bind        1  with-allocator
 7  signal              1  defers on the transfer path

Forty programs refused by name rather than miscompiled, and one line of build output says which node and where. A backend that quietly omitted the guard would have compiled all forty and been wrong in a way no test distinguishes from a race.

What this does not do is measure what conditions cost. That is still unknown, and it is still the only row of item 15's table with nothing behind it. What is now known is the shape of the bill: the guard is per call site, the pad is per restart-case activation, fdefers needs a second exit path that no form in body can reach, and any function with fdefers at all is refused today — which is most of the prelude's file and container code, and is why the programs that use a Vec are not in the 40.

The honest no-plan bucket

Everything below is refused by name at build time, not silently wrong.

• Conditions, entire — the guard, the landing pad, emit_restart_case, emit_with_alloc, the transfer exit, and fdefers on it. Several hundred lines of emit.ml reimplemented from spec-conditions.md rather than ported.
• Bounds checks, which are the same work: check_at and check_slice cannot exist without the guard.
• Rt with an aggregate return, and with it most of the container runtime; Vec, Map and Pool have not been exercised at all.
• Fnval's indirection cell. FnAddr (Fnval n) emits the symbol, which is correct for a whole-program build and wrong the instant anything is redefined into it. This backend has no cells and no --dev; that is a deliberate restriction and not an oversight, but it is exactly item 15's question 5 waiting where it was left.
• f64 → i64 out of range, and INT64_MIN / -1. idiv raises SIGFPE where LLVM says undefined, and cvttsd2si answers the integer-indefinite value. Unchanged from item 15: these want a language decision, not a backend.
• Debug information. None. --x86 and --debug together are refused.
• Code size and speed. Not measured. Every value is in memory, every intermediate is a frame temporary, and a block copy is rep movsb; that is the trade the brief asks for and nobody has put a number on it.

The verdict

The wiring is done and it was the easy half. What is left is conditions, and the measurement moved them from "first obstacle" to "the only obstacle".

The order item 15 recommended was floats first and conditions last. Floats turned out to be one afternoon's encodings and they are done. Conditions are still last and are now the whole remainder: they are what stands between 41 programs and the corpus, they are what a bounds check is made of, and check_no_transfer is the line that says so out loud on every build until someone writes them.

Two things are worth doing before that, and both are cheap. Decide what a bounds violation means in a build with no handler — because "no check" is what it means today and nothing says so. And take item 15's question 4 seriously now that there are two backends to disagree: (uninit) and unreachable already differ, deliberately, and the difference is currently documented only in a comment in x86.ml.