sx/current/CHECKPOINT-COMPILER-API.md

# CHECKPOINT-COMPILER-API — comptime `compiler` library (`#library "compiler"` + `abi(.zig) extern`)

Companion to the design-of-record
[../design/comptime-compiler-api.md](../design/comptime-compiler-api.md) (the plan
+ phased build order live there). This stream supersedes the metatype
`declare`/`define`/`type_info` `#builtin`s and the `#compiler` struct attribute
with ONE welded mechanism. Branch: `reify` (off `master`). Update after every step.

## ⏯ Resume (fresh session)

> **⚠ DIRECTION CHANGED (2026-06-17). The active plan is now
> [`PLAN-COMPILER-VM.md`](PLAN-COMPILER-VM.md), NOT the weld.**
> The **byte-weld + serialization/marshaling** approach is the wrong direction and is
> being **stripped**. New foundation: a **bytecode VM over flat, byte-addressable
> memory** so comptime values are native bytes; then the compiler-API rides on it with
> direct memory access (no weld, no validation, no marshaling). Everything below this
> banner describes the now-superseded weld state (committed on `reify` through
> `40d075c`) and is kept only to scope the Phase 0 strip. Read
> `PLAN-COMPILER-VM.md` first.
>
> **Why the pivot:** the comptime evaluator (`src/ir/interp.zig`) represents values as
> tagged `Value` unions, NOT native bytes — so a comptime `@ptrCast(*StructInfo)`
> reads the `Value` union's memory, not a struct. The weld tried to bridge that with
> hand-marshaling — exactly what the design set out to kill. Flat memory makes comptime
> values real bytes, so the bridge disappears. (JIT-native comptime was rejected: it
> breaks cross-compilation — host vs target layout — and loses the sandbox. A
> flat-memory VM keeps both while getting native bytes + speed.)
>
> **Next action (2026-06-18) — VM-native metatype CONSTRUCTION landed (step 7, uncommitted).** The
> metatype `declare`/`define` builtins + tagged-union `enum_init`-with-payload now run NATIVELY on
> the VM (new `.call_builtin` exec arm → `callBuiltinVm`/`defineFromInfo`, reading the `TypeInfo`
> value from FLAT MEMORY; faithful port of legacy `defineEnum`/`Struct`/`Tuple`). So the metatype
> CONSTRUCTION examples run **fully HANDLED** on the VM (no `call_builtin` fallback): `0614`/`0620`/
> `0621`/`0624`/`0632`; `0622`/`0623` define-HANDLED then fall back at the still-unported `type_info`.
> Both `enum_init`/`define` bail loudly on a `backing_type` tagged union (wrong layout) rather than
> silent-clobber. **697/0 BOTH gates + all unit tests** (added: tagged-union `enum_init` payload).
> NOT yet committed. **THE NEXT STEP: port `type_info`** (reflect a type → build the `TypeInfo`
> value in flat memory, the inverse of `define` — reuses tagged-union `enum_init`) so `0619`/`0622`/
> `0623` go fully HANDLED; then drive the SX_COMPTIME_FLAT_TRACE fallback list toward
> genuinely-non-comptime cases. Earlier landed: dedicated `Type` builtin TypeId (`6844fb9`/`94f60c5`/
> `554871b`); WRITE side declare_type/register_type/pointer_to VM-native (`66005af`); real
> lowering-time Context for allocating type-fns (`eb68d9e`). What's LEFT toward the end-state ONE
> evaluator: (1) finish porting the comptime corpus onto the VM (type_info next); (2) THEN flip the
> VM to default + delete `interp.zig` (with user go-ahead); (3) re-express `define`/`make_enum` as
> sx over the compiler-API once legacy is gone (allocation works only on the sole VM evaluator).
>
> Done so far in Phase 3:
> - **READ side (7 readers, dual-path):** `find_type`/`type_kind`/`type_field_count`/
>   `type_nominal_name`/`type_field_name`/`type_field_type`/`type_field_value`, each backed by a
>   `TypeTable` query both the legacy handler and the VM call (no drift). Examples 0628–0630.
> - **WRITE side (P3.3, legacy-only at lowering time):** `declare_type` + `pointer_to` + ONE
>   kind-branching `register_type` (subsumes `define`'s per-kind dispatch; codes match
>   `type_kind`: 1 struct · 2 actual `.@"enum"` · 3 tagged_union · 4 tuple). Idempotent re-fill
>   (two-edge import). Plus two fixes (issue 0142): all-void enum → real `.@"enum"` (was a
>   verifySizes panic); bare `EnumType.variant` qualified construction. Examples 0631–0635, 0187.
> - **Lowering-time VM (P3.4):** hardened the VM against malformed lowering-time IR (`refTy`,
>   bailing `aggType`, bounds-checked branch targets — bails, never panics); wired `tryEval`
>   into `runComptimeTypeFunc` behind the flag with legacy fallback; materialized a zeroed
>   lowering-time `Context` (the global isn't built yet at lowering). All measured green.
>
> **THE WALL (next step):** a `Type` *value* is an 8-byte tid, but `.any` (the boxed-any) is a
> 16-byte `{tag,value}` — and they share one TypeId (`.any`). So a `Type` in an aggregate
> (`Member.ty`/`EnumVariant.payload`) is sized 16B while the value is 8B → every lowering-time
> type-fn bails at `const_type` / the Member-array build. Can't make `kindOf(.any)` a word:
> at EMIT time `.any` really is a 16B box (variadic any, 0603), so that would silently corrupt
> it. The correct fix is a **dedicated `Type` builtin TypeId (8B), distinct from `.any`** —
> measured at **~123 `.any` references across ~25 files** (pack.zig has 30), a ~100-touch-point
> cross-cutting change → its own focused session (USER CHOSE to pause rather than rush it).
> Rejected alternatives: a scoped "lowering-mode treats `.any` as a word" flag (silent-wrong on
> a real Any box in a reflection type-fn); scalar-only Type-fns (safe but no real corpus type-fn
> is scalar-only — they all build a Member/variant aggregate).
>
> **Decisions recorded:** `find_type` returns a non-optional `TypeId` using `unresolved`(0), NOT
> `?Type`; reader names use the `type_*` family (avoid colliding with std `field_name`/`type_name`);
> the write side is a single kind-branching `register_type`; the write side stays LEGACY-only
> until the VM runs at lowering time (needs the `Type` TypeId). End-state guarantee: ONE
> evaluator — `interp.zig` deleted; dual-path + fallback are transitional (see PLAN end state).
> Build/verify: `zig build && zig build test` (**697**, gate OFF). Run the corpus ON the VM:
> `zig build test -Dcomptime-flat` OR env `SX_COMPTIME_FLAT=1`. Coverage trace:
> `SX_COMPTIME_FLAT_TRACE=1` (now also prints lowering-time `type-fn` HANDLED/fallback lines).

### (superseded) prior weld resume
Phase 1 done; Phase 2 welded structs were working via reflection + memory-order
validation (the `computeWeldPlan`/byte-blob "GEP engine" was explored + DROPPED even
earlier). A welded `Name :: struct abi(.zig) extern compiler { … }` declared fields in
the compiler type's MEMORY order; the compiler reflected the bound Zig type and
VALIDATED the header. **This whole mechanism is now being stripped — see the banner.**

> ⚠ Snapshot workflow: use `-Dname=examples/NNNN-foo.sx[,…] -Dupdate-goldens` to
> regenerate ONLY the named example(s) — a full `-Dupdate-goldens` re-runs all ~690
> and a flaky/host-divergent example (AOT/cross-arch) can clobber good snapshots.
> See CLAUDE.md → Snapshot integrity.

## Last completed step
**Phase 2 — welded structs by reflection + memory-order validation (byte-identical,
no GEP engine).** A welded `struct abi(.zig) extern compiler { … }` now works
end-to-end as a byte-identical mirror of the bound Zig type.

Design (locked, supersedes the byte-layout-override plan):
- The sx header declares fields in the compiler type's MEMORY order. The compiler
  REFLECTS the bound Zig type — field names from `@typeInfo`, offsets from
  `@offsetOf`, size from `@sizeOf` — and validates the header matches. Nothing is
  maintained by hand; a `types.zig` change re-reflects on the next compiler build.
- On pass it's an ORDINARY struct whose natural layout already equals the Zig
  layout → `@ptrCast` to the compiler type + deref is byte-identical. No
  byte-blob, no index/remap tables, no reorder, no special LLVM path.
- Loud, precise diagnostics on any drift: *field not found* (+ memory order),
  *wrong field order at position N* (+ expected memory order), *type layout
  mismatch* (field size), *layout mismatch* (total size / count).

What changed from the dropped plan:
- `compiler_lib.zig`: `weldStruct` now REFLECTS field names (`@typeInfo`) and bakes
  `bound_types` fields in ascending-OFFSET (memory) order — no hand-listed names.
  Deleted `computeWeldPlan`/`WeldPlan`/`WeldElement`. `validateStructLayout` checks
  the sx header against the memory-ordered registry.
- `nominal.zig` `validateWeldedStruct`: renders the precise diagnostics
  (+ `weldedFieldOrderStr`).
- Examples: `0627` (StructInfo in memory order, byte-identical, usable);
  `1186` (source-order StructInfo → wrong-field-order diagnostic). `1183` message
  refreshed.
- `zig build` + `zig build test` green (692 corpus, unit tests pass).

### Earlier — Phase 2.1 (weld-plan layout math, now removed)
**The weld-plan offset math + `StructInfo` registered.** Was the core of the
byte-layout-override engine; superseded by the reflection+validation design above.

Decision (locked 2026-06-17): **full byte-layout weld** — a welded sx struct is
laid out byte-identically to the bound Zig type (Zig's `@offsetOf`, reordering +
padding included), so it passes to a Zig handler as raw memory with zero
marshalling. (The alternative — handlers reading interp `Value` aggregates
logically, no layout override — was rejected; welded types must also be usable as
runtime data, and the design wants the literal byte weld.)

- Measured: Zig reorders `StructInfo` to `fields`@0, `name`@16, `nominal_id`@20,
  `is_protocol`@24, size 32 — vs sx-natural `name`@0, `fields`@8, … So the override
  is genuinely required (`Field`'s two-u32 natural layout was the easy case).
- `compiler_lib.zig`: registered `StructInfo` (`weldStruct`, the second
  `bound_types` entry). Added `WeldElement` / `WeldPlan` + `computeWeldPlan(alloc,
  fields, total)` — pure: orders fields by ascending byte offset, inserts padding
  elements for gaps + the alignment tail, and builds the sx-field → LLVM-element
  remap. This is what the LLVM type builder + struct-GEP sites will consume.
- Unit-tested (`compiler_lib.test.zig`): `Field` → identity plan (2 elems, no pad);
  `StructInfo` → 5 elems `[fields@0, name@16, nominal_id@20, is_protocol@24,
  pad@25..32]`, remap `[1,0,3,2]`.
- `zig build` + `zig build test` green.

### Earlier — Phase 1 polish (comptime-only enforcement)
**A RUNTIME call to a `fn abi(.zig) extern compiler` is a clean build-gating error
instead of an undefined-symbol link failure.**
- `emitCall` (`src/backend/llvm/ops.zig`): when the callee is `compiler_welded`
  AND the ENCLOSING function is not `is_comptime` (i.e. genuine runtime code, not a
  `#run`/`::` initializer wrapper whose LLVM body is dead), print a clear
  "comptime-only … cannot be called at runtime" error and set
  `comptime_failed` (the driver halts before object/JIT emission). The enclosing
  `is_comptime` guard is what keeps the legitimate `#run` use (example 0626) green.
- Corpus: `examples/1185-diagnostics-weld-fn-runtime-call.sx` (runtime `intern(…)`
  → clean error, exit 1, no link failure).
- `zig build` + `zig build test` green (458 unit + 690 corpus).

### Earlier — fifth sub-step (host-call bridge)
**A `fn abi(.zig) extern compiler` dispatches, under the comptime interpreter, to
its registered Zig handler instead of dlsym.**
- `compiler_lib.zig`: function registry — `BoundFn { sx_name, handler }`,
  `bound_fns` = `intern(string)->StringId` + `text_of(StringId)->string` (the
  string-pool round-trip), `findFn`, and `FnHandler` (`*Interpreter, []Value ->
  Value`). `intern` mutates via `interp.mint orelse @constCast(&module.types)`
  (the same mutable-table access the metatype mint path uses); `text_of` reads the
  const pool. Imports `interp.zig` (the compiler_hooks↔interp cycle pattern).
- IR `Function` gained `compiler_welded: bool`. `declareFunction`
  (`src/ir/lower/decl.zig`) sets it via `weldedCompilerFn`, which also VALIDATES:
  the bound lib must be `compiler` and the name must be on the function-export
  list — else a build-gating `.err` (no silent fall-through to dlsym).
- `interp.call()`: before the dlsym/extern path, a `compiler_welded` function
  routes to `compiler_lib.findFn(name).handler(self, args)` (clean bail off the
  export list).
- Corpus: `examples/0626-comptime-weld-fn-intern-text-of.sx` (`#run
  text_of(intern("hello, compiler"))` folds to a string constant → prints it);
  `examples/1184-diagnostics-weld-fn-unexported.sx` (unexported welded-fn name →
  build error). `findFn` lookup unit-tested.
- **Runtime-call rejection is NOT yet clean** — welded fns are comptime-only; a
  RUNTIME call would emit a reference to a non-existent extern symbol → a loud
  LINK error (not silent, but not a tidy diagnostic). The examples call welded fns
  only inside `#run`. A dedicated "comptime-only symbol" emit diagnostic is the
  immediate follow-up.
- `zig build` + `zig build test` green (458 unit tests + 689 corpus).

### Earlier — fourth sub-step (welded-struct layout validation)
**A `struct abi(.zig) extern compiler { … }` is validated against the binding
registry as a *header checked against the implementation*.**
- `compiler_lib.zig`: `validateStructLayout(bt, sx_fields, total)` — pure, returns
  the first `LayoutMismatch` (field count / name / size / total) or null. Plus
  `lib_name = "compiler"` and `SxField`. Unit-tested (faithful `Field` passes;
  each drift flagged as the right variant).
- `registerStructDecl` (`src/ir/lower/nominal.zig`): for `sd.abi == .zig`,
  `validateWeldedStruct` checks the bound lib is `compiler`, the name is on the
  export list (`findType`), and the sx layout (field names + `typeSizeBytes` +
  total) matches the welded type — emitting a build-gating `.err` (good span into
  the struct body) on any failure. No silent reinterpretation.
- `#library "compiler"` is the comptime-only internal surface, NOT a dylib —
  `src/main.zig`'s dlopen walker skips it (was emitting a spurious `libcompiler.so`
  load warning).
- Corpus: `examples/0625-comptime-weld-struct-field.sx` (faithful `Field` welds,
  validates, usable as data → `name=7 ty=3`); `examples/1183-diagnostics-weld-
  struct-field-count.sx` (one-field `Field` → build-gating field-count diagnostic).
- **Offset-override / GEP emission for non-natural Zig layouts is NOT here** — it
  isn't exercised by `Field` (two u32s = natural layout coincides with the weld).
  It arrives with `StructInfo` in Phase 2 (slices/reordering), where the bound
  offsets actually differ from the sx-natural ones. The validation already checks
  per-field size + total, so a layout drift is caught even before the override
  engine exists.
- `zig build` + `zig build test` green (456 unit tests + 687 corpus).

### Earlier — third sub-step (binding registry)
**The binding registry (welded-type lookup, layout baked from the real Zig
type).**
- New `src/ir/compiler_lib.zig` — the `compiler` library's binding registry, the
  curated safety boundary. `BoundType { sx_name, size, alignment, fields:
  []FieldLayout{name, offset, size} }`; `weldStruct` bakes the layout from a real
  Zig struct via `@sizeOf`/`@alignOf`/`@offsetOf` at compiler-build time (a
  sx-field-count mismatch is a `@compileError`, never a silent truncation).
  `bound_types` exports `Field` (welded to `types.TypeInfo.StructInfo.Field` —
  two `u32`s); `findType(sx_name) ?*const BoundType` is the lookup the welded-decl
  resolution path will consult (returns null off the export list — clean boundary,
  no silent default).
- Registered in the barrel (`src/ir/ir.zig`): `compiler_lib` + `compiler_lib_tests`.
- Tests (`src/ir/compiler_lib.test.zig`): `findType("Field")` equals the real
  `StructInfo.Field` `@sizeOf`/`@alignOf`/`@offsetOf` (8 bytes, two u32s at 0/4);
  an unexported name returns null. Break-verified (a wrong size → suite red,
  named `ir.compiler_lib.test...`).
- `zig build` + `zig build test` green (454 unit tests).

### Earlier — second sub-step (struct-decl parse)
**`abi(.zig) extern <lib>` PARSES on a STRUCT decl (parse-only, no semantics).**
- `ast.StructDecl` gained `abi: ABI` + `extern_lib: ?[]const u8` binding fields.
- `parseStructDecl` (`src/parser.zig`): after `struct` (and the `#compiler`
  check), parse an optional `abi(...)` then optional `extern <lib>` — same slot
  order as fn decls — and thread them onto the node. Ordinary structs are
  unperturbed (`parseOptionalAbi`/`parseOptionalExternExport` no-op when absent).
- Parser unit tests (`src/parser.test.zig`): `Field :: struct abi(.zig) extern
  compiler { name: StringId; ty: Type; }` parses with `abi == .zig`, `extern_lib
  == "compiler"`, field list intact; a plain struct leaves `abi == .default` /
  `extern_lib == null`. Break-verified (a wrong-sentinel assert turns the suite
  red, confirming the test runs).
- `zig build` + `zig build test` green.

### Earlier — first sub-step (fn decls) + the syntax pivot
**`abi(.zig) extern <lib>` PARSES on a fn decl (parse-only).** Plus the syntax
pivot it required.

Syntax decision (locked 2026-06-17, supersedes the doc's original
`extern(.zig) <lib>` single-qualifier form): the ABI/layout selector and the
linkage keyword are two orthogonal annotations.
- `abi(.x)` — ABI / calling-convention annotation in the slot **before**
  `extern`/`export`. **Unified replacement for `callconv(...)`, which is removed.**
  `ABI = { default, c, zig, pure }`: `.c` (C ABI), `.zig` (Zig-layout weld → the
  `compiler` library), `.pure` (naked asm), `.default` (unannotated). Can appear
  standalone (no extern) on any fn / fn-type / lambda.
- `extern <lib>` — linkage keyword + binding source (named library).

So a welded binding is `text_of :: (id: StringId) -> string abi(.zig) extern compiler;`.

What landed:
- **AST** (`src/ast.zig`): `CallingConvention` → `ABI { default, c, zig, pure }`;
  the `call_conv` field → `abi: ABI` on `FnDecl` / `Lambda` / `FunctionTypeExpr`.
- **Lexer/token** (`src/token.zig`, `src/lexer.zig`): `kw_callconv` → `kw_abi`,
  keyword string `"callconv"` → `"abi"`.
- **Parser** (`src/parser.zig`): `parseOptionalCallConv` → `parseOptionalAbi`
  (parses `abi(.c|.zig|.pure)`); wired in the fn-decl postfix slot (before
  `extern`/`export`), the function-type-expr slot, and the lambda slot;
  `isFunctionDef`/`hasFnBodyAfterArrow` recognise `kw_abi`.
- **AST→IR map** (`src/ir/type_resolver.zig`, `src/ir/lower/decl.zig`, `sema.zig`,
  `closure.zig`): the AST `.abi == .c` reads kept their C-ABI meaning; the
  function-type resolver maps `.zig`/`.pure` → IR `.default` (no fn-pointer-type
  CC for those decl-level ABIs; neither occurs in a function-TYPE position yet).
- **CC-mismatch diagnostic** (`src/ir/lower/expr.zig`, `src/sema.zig`): the
  user-facing text `callconv(.c)` → `abi(.c)`.
- **sx migration**: 52 `.sx` files `callconv(` → `abi(` (all were function-type
  callback annotations — none in the fn-decl postfix slot, so no reordering).
- **Docs**: `readme.md`, `specs.md`, the design doc, snapshots (0114 / 1104 /
  1200) regenerated for the rename.
- **Tests**: parser unit tests in `src/parser.test.zig` — `abi(.zig) extern <lib>`
  on a fn decl (asserts `abi == .zig`, `extern_export == .extern_`, `extern_lib ==
  "compiler"`); bare `extern` leaves `abi == .default`; standalone `abi(.c)` /
  `abi(.pure)`. lexer/sema tests updated.

`zig build` + `zig build test` green (450/450 unit + 685 corpus).

## Current state

> **Pivoted — see the banner + `PLAN-COMPILER-VM.md`.** The items below are the weld
> machinery as it stands on `reify` HEAD (`40d075c`); they are the **strip list** for
> Phase 0, not the forward direction. The `#library`/`abi`/`extern` *syntax* stays; the
> weld *semantics* (layout reflection/validation, marshaling dispatch) go.

- `compiler :: #library "compiler";` parses + is recognised as the comptime-only
  internal surface (never dlopen'd).
- `abi(.zig) extern compiler` STRUCTS: layout-validated against the registry
  (faithful → ok; drift → build-gating diagnostic). `Field` welds + usable.
- `abi(.zig) extern compiler` FUNCTIONS: dispatched under the comptime interp to
  their registered Zig handler (`intern`/`text_of` round-trip works); unexported
  names rejected at declaration. Comptime-only.
- A RUNTIME call to a welded fn is a clean build-gating error (comptime-only
  enforcement at `emitCall`); the legitimate `#run`/`::` use stays green.
- The whole Phase 1 foundation (parse → registry → struct-layout validation →
  function host-call bridge → comptime-only enforcement) is in place for the
  two-u32 `Field` case + the two string readers.
- **Deferred**: offset-override / LLVM byte-offset GEP for non-natural layouts
  (needed by `StructInfo`'s slice field, Phase 2).

## Next step — execute `PLAN-COMPILER-VM.md`

> The weld is being stripped. The next step is **Phase 0 of
> [`PLAN-COMPILER-VM.md`](PLAN-COMPILER-VM.md)** — remove the weld / serialize /
> marshal machinery (`compiler_lib.zig` reflection+validation, `nominal.zig`
> `validateWeldedStruct`, the `compiler_welded` dispatch, the weld examples/diagnostics
> 0625/0627/1183/1184/1185/1186), keeping the `#library`/`abi`/`extern` *syntax*. Then
> Phase 1 (flat-memory value model). The weld-era "next step" below is **obsolete** —
> kept only as a record of what the weld surface was about to do.

### (obsolete) weld-era next step
Welded structs were byte-identical mirrors, so the API surface was set to grow:

- **Bind `register_struct` / `find_type`** over the host-call bridge
  (`compiler_lib.zig` `bound_fns`, like `intern`/`text_of`). `register_struct`
  takes a welded `StructInfo` and mints a real `TypeId` (guarded: dup field names,
  kind well-formedness — the checks `define` does today). Because the welded
  `StructInfo` is byte-identical, the handler can read it as the real Zig
  `*StructInfo` (cast + deref) rather than marshalling a `Value` field-by-field —
  the payoff of the byte-weld. `find_type(StringId) -> ?Type` reads the table.
  Prove: build a struct programmatically + round-trip a source one.
- **Re-express `type_info`/`define` (struct) as sx** over `register_struct`/
  `find_type`; migrate `examples/0622`; delete the bespoke struct interp arms
  (`defineStruct` / the `reflectTypeInfo` struct path).

Then Phase 3+: widen the welded types to `EnumInfo`/`TaggedUnionInfo`/`TupleInfo`
(optional fields → sentinels) — each just needs an sx header in the compiler
type's memory order + the matching `register_*` fn. Finally migrate `BuildOptions`
to `abi(.zig) extern compiler` (re-home the `#compiler` registry) and delete
`#compiler`.

Note: a welded struct with an `?T` / `union(enum)` field (e.g. `EnumInfo`'s
`backing_type: ?TypeId`, `explicit_values: ?[]const i64`) is the next layout
wrinkle — the sx header must mirror Zig's optional/union representation. Handle
when reached (sentinels or accessor fns; see the design doc Risks).

## Known issues
- None for this stream. (Metatype's deferred enhancement is issue 0141 — comptime
  `List` growth; orthogonal, see `current/CHECKPOINT-METATYPE.md`.)

## Log
- **Phase 3 P3.4 step 7 (VM plan) — VM-native metatype CONSTRUCTION: `declare`/`define` + tagged-union `enum_init` (2026-06-18).**
  Ported the metatype type-CONSTRUCTION builtins into the VM so the construction examples run
  HANDLED end-to-end (no `call_builtin` fallback). Three pieces: (1) **tagged-union `enum_init`
  with payload** — the arm previously bailed; now allocates the value (zeroed), writes the tag at
  offset 0 (`{ header(tag)@0, [N x i8] payload@tag_size }`, the LLVM `backend/llvm/types.zig`
  layout) and copies the payload at `tag_size`. (2) A **`.call_builtin` exec arm** → new
  `callBuiltinVm`, the VM-native mirror of the legacy `execBuiltinInner`: `declare(name)` mints an
  empty forward nominal slot (shared `declareNominal` helper, also used by `declare_type`);
  `define(handle, info)` reads the `TypeInfo` tagged-union VALUE from FLAT MEMORY (tag@0, active
  payload `EnumInfo`/`StructInfo`/`TupleInfo` struct at `tag_size`, its single slice field) and
  mints via `defineFromInfo`, a faithful port of legacy `defineEnum`/`defineStruct`/`defineTuple`
  (all-void enum → real `.@"enum"` per issue 0142, dup-name rejection, `updatePreservingKey` vs
  `replaceKeyedInfo`). (3) Refactored the `[]{name,ty}` decode out of `registerTypeVm` into a
  shared `decodeMemberSlice` (+ `decodeTypeSlice` for bare-`Type` tuple elements), keyed to the
  module-level `NamedMember`. Unmodeled builtins (`type_info`/`type_name`/…) return null → bail
  with the builtin name → legacy fallback (dual-path parity). **Correctness guard (caught via
  review):** `enum_init`/`define` assume a tag-headed layout, which is WRONG for a `backing_type`
  tagged union (laid out as the backing struct) — both now bail loudly on `backing_type != null`
  rather than silent-clobber. **Result:** examples `0614`/`0620`/`0621`/`0624`/`0632` run **fully
  HANDLED** on the VM (define is the whole eval); `0622`/`0623` run define HANDLED then fall back
  cleanly at the still-unported `type_info` reflection. VM output byte-matches legacy for all 7.
  **697/0 BOTH gates + all unit tests (added: tagged-union `enum_init` payload layout).** On
  `reify`. **Next:** port `type_info` (REFLECT a type → build a `TypeInfo` value in flat memory,
  the inverse — reuses the tagged-union `enum_init` write) so `0619`/`0622`/`0623` go fully HANDLED;
  then the rest of the comptime corpus (drive the SX_COMPTIME_FLAT_TRACE fallback list toward the
  genuinely-non-comptime cases) before the VM-default flip + legacy deletion.
- **Phase 3 P3.4 step 6 (VM plan) — REAL lowering-time Context: allocating + List-building type-fns now run HANDLED on the VM (2026-06-18).**
  The VM can now evaluate a comptime type-fn that ALLOCATES at lowering time (the 0141 family) —
  the legacy interp cannot. Four changes: (1) `runComptimeTypeFunc` (lower/comptime.zig) FORCES the
  CAllocator→Allocator thunks to exist (`getOrCreateThunks`, idempotent, guarded by Allocator/
  CAllocator registered) BEFORE eval — a type-fn const runs at scanDecls (Pass 1), before Pass 1c
  builds the default-context global + thunks, so the comptime allocator was otherwise null;
  (2) `materializeDefaultContext` builds a REAL context at lowering time when the global is absent —
  finds the two thunks by name (`findFuncByName`) and lays their func-refs into the inline
  `Allocator` value `{ctx=null, alloc_fn@+ptr, dealloc_fn@+2*ptr}` at the head of `Context`, so
  `context.allocator.alloc_bytes` dispatches `call_indirect` → thunk → native VM `malloc`;
  (3) `aggType` now DEREFS a pointer `base_type` (the List write path emits `struct_gep` with
  `base_type = *Struct` — `fieldOffset` panicked on the pointer; now derefs to the pointee, no
  panic); (4) `subslice` handles a `[*]T` many-pointer / `*T` base (a List's `items` field — the
  base IS the data pointer). **Verified end-to-end (manual probe):** a compiler-API type-fn that
  builds its `[]Member` in a `List(Member)` (`.append` ×3, then `register_type(handle, kind,
  vs.items[0..vs.len])`) runs **HANDLED on the VM** and mints correctly (`green=7`) — the exact
  0141 List-growth pattern, on the VM. **Can't be a corpus test yet** (gate-OFF/legacy still can't
  allocate at lowering time — the dual-path bind), so locked in via VM unit tests instead
  (many-pointer subslice; `struct_gep` with a pointer `base_type`). **697/0 BOTH gates + all unit
  tests, EXIT=0.** On `reify`. **Remaining for the original 0141 repro (uses metatype `define`/
  `make_enum` → `call_builtin` → legacy fallback → legacy fails):** re-express the metatype over the
  compiler-API so the whole type-fn runs on the VM (no `call_builtin`). THEN the repro works on the
  VM — and the dual-path bind resolves only at the VM-default-flip + legacy-deletion end-state.
- **Phase 3 P3.4 — investigation: the "real lowering-time Context" is BLOCKED by issue 0141 (2026-06-18).**
  Probed whether the VM needs a REAL lowering-time `Context` (CAllocator thunk func-refs) for
  allocating type-fns. **Finding: lowering-time comptime ALLOCATION fails in the LEGACY interp
  too** — a type-fn that calls `context.allocator.alloc_bytes` at lowering time bails in legacy
  with `comptime call_indirect: callee is not a func_ref Value (raw fn-pointers from extern calls
  aren't dispatchable in interp)`, and the VM bails at parity (`call_indirect through a null
  function pointer`). This is exactly issue **0141**'s root cause (its analysis already notes "the
  null allocator is the same story for the CAllocator thunks") — an OPEN deferred issue. So:
  (1) the VM is CORRECT (parity — both bail; no regression); (2) the real-context work is
  PREMATURE — its only consumer (allocating lowering-time type-fns) can't pass gate-OFF, so no
  corpus test can validate it, and even a more-capable VM can't ship a divergence during the
  dual-path phase. **Consequence for the metatype re-expression:** re-expressing `define`/`make_enum`
  over the compiler-API needs to BUILD `[]Member` slices dynamically (allocation) — which is
  blocked by 0141 at lowering time. The viable paths are: (a) avoid allocation by passing the
  caller's existing slice through (needs `EnumVariant`/`StructField` to be usable AS `Member` —
  they're layout-identical `{string, Type}`, but distinct nominal types — a metatype-API decision),
  or (b) wait for 0141. **No code change this step** (the VM already bails correctly). Recorded so
  the next session doesn't re-derive it. 697/0 both gates unchanged.
- **Phase 3 P3.4 step 5 (VM plan) — WRITE side ported to the VM → FIRST HANDLED lowering-time type-fns (2026-06-18).**
  Ported `declare_type` / `pointer_to` / `register_type` into `Vm.callCompilerFn`, mirroring the
  legacy `compiler_lib` handlers (mint via `@constCast(table)` — the same mutable access the
  read-side `intern` uses; the lowering-time mint target IS `&module.types`). `register_type`
  reads the `[]Member` slice from FLAT MEMORY: threaded `ref_types` through `invoke` →
  `callCompilerFn` so the slice's element type (`Member = {name: string, ty: Type}`) gives the
  field offsets + stride; decodes each `{name, ty}` and branches on `kind` (1 struct · 2 enum ·
  3 tagged_union · 4 tuple) exactly as legacy (dup-name / payload-on-enum rejections, idempotent
  re-fill via `nominalIdentOf`). **Key unblock:** the synthesized comptime type-fn wrapper
  (`createComptimeFunction`/`…WithPrelude`) was built with return type `.any` → `regToValue`
  bailed at the VM↔legacy boundary; changed to `.type_value` (the legacy path reads via `asTypeId`
  regardless, so no legacy change). **Result: the compiler-API write type-fns now run HANDLED
  end-to-end on the VM at LOWERING time** — `0631` (register-graph: 2 HANDLED, A↔B cycle via
  forward handles + `pointer_to`) and `0635` (multi-edge import: 2 HANDLED), parity-correct. They
  run on the ZEROED lowering-time context (fixed `.[…]` member arrays, no allocation). The
  metatype `make_enum`/`define` examples (`0632`) still fall back CLEANLY through
  `call_builtin(define)` (the separate metatype path — re-expressing it onto the compiler-API is
  the other half of P3.4). **697/0 BOTH gates + EXIT=0.** On `reify`. **Next:** (optional, deferred)
  a REAL lowering-time Context (CAllocator thunk func-refs) for List-growing type-fns; and
  re-express the metatype `define`/`make_enum` over the compiler-API to delete the bespoke interp
  arms (the end-state: ONE evaluator).
- **Phase 3 P3.4 step 4 (VM plan) — model `.type_value` natively in the comptime VM (2026-06-18).**
  The VM now HANDLES Type values instead of bailing: `kindOf(.type_value)` → `.word`; a new
  `const_type` exec arm → the word `TypeId.index()`; `regToValue` maps a `.type_value` word back
  to a `.type_tag` Value at the legacy boundary (`valueToReg` already mapped `.type_tag` →
  index). Surfaced + fixed a VM PANIC (forbidden): `struct_init` assumed a `.@"struct"` result
  type and union-access-panicked on an ARRAY literal (`EnumVariant.[ … ]`, reached now that Type
  args no longer bail early) — it's the generic aggregate-literal op, so it now dispatches on the
  result kind (struct / array / tuple) and BAILS loudly on anything else, never panics. **697/0
  both gates** (the make_enum type-fns now run further on the VM, then bail cleanly at the
  `define`/`make_enum` `call_builtin` → legacy mints — no mutation before the bail, parity holds).
  VM unit test added (const_type → word → regToValue → `.type_tag`). On `reify`. **Next (the
  payoff):** port the WRITE side (declare_type / register_type / pointer_to) into
  `Vm.callCompilerFn` + give the lowering-time path a REAL Context (CAllocator thunk func-refs,
  not zeroed) → the first HANDLED lowering-time type-fn end-to-end on the VM.
- **Phase 3 P3.4 step 3 (VM plan) — dedicated `Type` builtin TypeId: RESOLVER FLIPPED + `.any` migration (2026-06-18).**
  Flipped `type_resolver:64` (`"Type"` → `.type_value`), `module.zig` `constType` (result type
  → `.type_value`), and `emitConstType` (a bare i64 carrying `tid.index()`, NOT a 16-byte Any
  box). Then migrated every `.any` reference that means "a Type value", classified per CLAUDE.md
  (leave the real boxed-Any refs): (a) the "Any holds a Type" **meta-marker tag** moved `.any` →
  `.type_value` at all 4 consumers — `reflectArgTypeId` (LLVM), `reflectTypeId` + the
  `.type_tag`-as-struct-field comptime path (interp), and `resolveTypeCategoryTags("type")`
  (generic.zig); (b) reflection-builtin RETURN types `.any` → `.type_value` (`type_of`/`declare`/
  `define`); the runtime `type_of(any)` now reads the tag AS a `.type_value` (no re-box); (c)
  expr_typer infers a bare type-name expr as `.type_value` (with a `is_raw` backtick exemption —
  `` `string `` is a value, never the reserved type); (d) `reflectionArgIsType` accepts
  `.type_value` OR `.any` (a reflection arg can be a bare Type OR a boxed Any — the over-narrow
  `==.type_value` was the catastrophic-regression cause, caught + fixed); (e) the comptime
  `switch_br` accepts a `.type_tag` discriminant (type-category match); (f) a bare function name
  in a `Type` slot now lowers to `const_type(its real function type)` instead of a func-ref
  (fixed a JIT crash — was a func-ref word read as a TypeId), keeping the old string-box path only
  for genuine `Any` params; (g) the field-not-found diagnostic + `formatTypeName` render
  `.type_value` as "Type". Fixed 3 unit tests asserting the old `.any` Type behavior.
  **697/0 BOTH gates** + all 494 unit tests (EXIT=0). Gate ON stays green because the VM's
  `kindOf(.type_value)` → `.unsupported` → bails CLEANLY to legacy (no silent-wrong) — the VM
  doesn't model `Type` values YET (next step), but parity holds. Regenerated 24 snapshots (22
  `.ir` const_type-shape; 2 `.stderr` Any→Type — diff reviewed, only the intended changes). On
  `reify`. **Next:** model `.type_value` natively in the VM (`kindOf` → word, `const_type` → word
  = `TypeId.index()`, `regToValue` word → `.type_tag`) for COVERAGE, then port the WRITE side into
  `callCompilerFn` + a real lowering-time Context → the first HANDLED lowering-time type-fn.
- **Phase 3 P3.4 step 2 (VM plan) — dedicated `Type` builtin TypeId: FOUNDATION landed (dead/additive) (2026-06-18).**
  Added `TypeId.type_value` (slot 19) + a matching `TypeInfo.type_value` variant + the builtins
  init entry — an **8-byte type handle distinct from the 16-byte boxed `.any`** (THE WALL). All
  `types.zig` layout handlers wired: `sizeOf`/`typeSizeBytes` → 8, `typeAlignBytes` → 8,
  `typeName` → "Type", `hashTypeInfo`/`typeInfoEql` no-payload arms. Only ONE exhaustive switch
  needed a new arm (`backend/llvm/types.zig` `toLLVMTypeInfo` → `cached_i64`); every other
  `switch(TypeInfo)` site has an `else` (audited when the resolver flips). **`first_user` 19 → 100**
  (per the user): slots 20–99 are RESERVED builtin headroom (infos padded with the `unresolved`
  tripwire), so future builtins don't renumber user TypeIds / churn `sx ir` snapshots. Cost:
  ~80 default entries in each binary's per-type reflection arrays (user opted in). **Still dead:**
  `type_resolver.zig:64` STILL returns `.any` for "Type" — nothing produces `.type_value` yet, so
  NO behavior change. Regenerated 22 IR snapshots (pure TypeId renumber to 100-base; `git diff
  --name-only` confirmed ONLY `.ir` files + the 2 source files changed — no stdout/stderr/exit).
  **697/0 both gates** (OFF and `-Dcomptime-flat`). **Next:** flip `type_resolver:64` →
  `.type_value`, then migrate the `.any` refs that mean "a Type value" (const_type result /
  reflection returns / metatype `Type` params / `.type_tag` checks) — leave the real boxed-Any
  refs — file-by-file with a build after each.
- **Phase 3 P3.4 step 1 (VM plan) — lowering-time default context; first blocker cleared (2026-06-18).**
  `materializeDefaultContext` now falls back to a ZEROED `Context` (found by name) when the
  `__sx_default_context` global is absent — i.e. at LOWERING time, where the global isn't
  emitted yet. A type-fn that never touches the allocator now runs past context setup; one
  that allocates reads a null `alloc_fn` (zeroed) → `call_indirect` on the null func-ref
  bails → legacy fallback (a REAL lowering-time context with the CAllocator thunk func-refs,
  so allocating type-fns also run on the VM, is a follow-up). **Measurement: the bail moved
  deeper** — metatype `make_enum` now bails at `const_type` (the `Type`-literal op, unported);
  `register_type` type-fns bail at the welded write call (declare_type/register_type aren't in
  `callCompilerFn`). No table mutation happens before either bail (the write fns bail before
  minting), so parity holds: both gates **697/0**, no crashes. **Next blockers (the "model
  Type" chunk):** (a) the `const_type` op → a word = `TypeId.index()`; (b) the Type-return
  bridge (`regToValue` for a `Type`/`.any` word → `.type_tag`); (c) the VM-native write side
  (declare_type/register_type/pointer_to in `callCompilerFn`) + a real lowering-time context.
  Only once those land does a type-fn actually run end-to-end on the VM (a HANDLED case).
- **Phase 3 P3.4 (VM plan) — wire the VM at the LOWERING-time site + measure (2026-06-18).**
  Routed `runComptimeTypeFunc` (the type-fn fold — the THIRD comptime call site) through
  `comptime_vm.tryEval` behind `-Dcomptime-flat`/`SX_COMPTIME_FLAT` with legacy fallback,
  mirroring the two emit-time folds. Extracted the shared post-check (`checkComptimeTypeResult`
  — the declared-but-never-defined zero-field guard) so both paths use it. **Measurement
  (SX_COMPTIME_FLAT_TRACE):** every metatype/compiler-API type-fn currently bails CLEANLY
  with `no __sx_default_context global to materialize the implicit context` — at lowering
  time the default-context global doesn't exist yet (it's built at emit time), so the VM bails
  at context materialization, BEFORE running the body (no partial mint, no crash → legacy
  mints). The hardening holds: **no crashes** across the corpus on the VM lowering-time path.
  Both gates **697/0**. **So the FIRST lowering-time blocker is the implicit context, not
  `Type` modeling** — the VM needs a way to materialize/skip the default context at lowering
  time (most type-fns get an implicit ctx for potential `List`-growth alloc; many don't use
  it). Next: materialize a lowering-time default context for the VM (or pass a null ctx +
  bail only if the allocator is actually used), THEN model `Type` values + the VM-native write
  side. This is near-pure fallback today — permanent scaffolding that lights up as those land.
- **Phase 3 P3.4-prep (VM plan) — harden the VM against malformed lowering-time IR (2026-06-18).**
  Prerequisite for wiring the VM at the LOWERING-time comptime site (`runComptimeTypeFunc`),
  where IR can be malformed (an unresolved name lowers to a dangling / `Ref.none` operand —
  the 0737 crash). Closed the remaining panic vectors so the VM BAILS (→ legacy fallback)
  instead of aborting: (1) a checked `Vm.refTy(ref_types, r)` replaces every raw
  `ref_types[ref.index()]` in `exec` (the type-side companion to `Frame.get`'s `bad_ref`
  value-side guard); (2) `aggType` is now a bailing method (`Error!TypeId`) using `refTy`;
  (3) the block-dispatch loop bounds-checks the branch target before indexing
  `func.blocks.items`. `global_get` was already guarded. No behavior change — gate OFF and
  ON both **697/0**; unit test added (a `cmp_lt` with a `Ref.none` operand bails, not
  panics). **Next:** wire `tryEval` into `runComptimeTypeFunc` behind the flag with legacy
  fallback and measure (most minting type-fns will still bail at the welded-write call /
  `Type`-result conversion until the VM models `Type` values + the VM-native write side land
  — those are the steps that actually move lowering-time comptime onto the VM, toward
  deleting legacy).
- **Phase 3 P3.3 (VM plan) — WRITE side: declare_type + pointer_to + ONE kind-branching register_type (2026-06-18).**
  The mutating compiler-API: `declare_type(name) -> Type` (forward handle), `pointer_to(t) -> Type`
  (build `*T`), and `register_type(handle, kind, members: []Member) -> Type` which branches on
  `kind` IN THE COMPILER (subsuming define's per-kind dispatch). Take/return real `Type` values
  (matching meta.sx declare/define). **Timing (per user): mint LAZILY at lowering time, single
  pass** (the existing `runComptimeTypeFunc`), so the write side is **legacy-only** (`compiler_lib`
  handlers) — the VM isn't wired at lowering time, no VM mirror needed; readers stay dual-path.
  A non-generic `-> Type` builder is now flagged `is_comptime` (decl.zig) so its dead body permits
  the welded calls. **Graph:** forward handles + `pointer_to` express mutually-recursive A↔B (`*A`,
  `*B`, B-by-value); `register_type` is **idempotent** (re-fill a nominal slot reached via two
  import edges — `nominalIdent`). `kind` codes match `type_kind` (1 struct · 2 actual `.@"enum"` ·
  3 tagged_union · 4 tuple). **Fixed two bugs (issue 0142):** (a) a fully payloadless minted enum
  was an all-void tagged_union → verifySizes panic; now a real `.@"enum"` (register_type kind 2 AND
  metatype `defineEnum`); (b) bare `EnumType.variant` payloadless qualified construction wasn't
  supported (failed for hand-written enums too) — added in `lowerFieldAccess` (`isPayloadlessVariant`).
  Examples 0631 (graph + actual enum + reflection), 0632 (make_enum all-void), 0633/0634/0635
  (namespaced / bare / multi-edge import of a minted type), 0187 (qualified variant construction).
  **Parity 697/697** (gate ON and OFF); unit tests added. **Next (P3.4):** re-express
  declare/define/type_info as sx over the compiler-API + delete the bespoke interp arms (needs the
  VM hardened for lowering-time IR, or the metatype migrated onto the legacy compiler-API calls).
- **Phase 3 P3.2b (VM plan) — kind + enum-value readers: `type_kind` + `type_field_value`; READ side complete (2026-06-18).**
  The last two read-only readers the metatype's `type_info(T)` needs (added to
  `compiler_lib.bound_fns` AND `Vm.callCompilerFn`, each backed by a `TypeTable` query both
  call): `type_kind(t) -> i64` (`kindCode` — a stable, compiler-owned discriminant: 0 other ·
  1 struct · 2 enum · 3 tagged_union · 4 tuple · 5 union · 6 array · 7 vector · 8 error_set;
  TOTAL, never bails) and `type_field_value(t, idx) -> i64` (`memberValue` — an enum variant's
  explicit value or ordinal; mirrors the `field_value_int` builtin; loud-bail for non-enum /
  out-of-range). Example `0630-comptime-compiler-type-kind` reflects `Color` / `WindowFlags`
  (flags) / `Point`. **The READ side is now COMPLETE** — `find_type` + `type_kind` +
  `type_field_count` + `type_field_name`/`type_field_type`/`type_nominal_name` +
  `type_field_value` cover everything `reflectTypeInfo` reads. VM unit test added. **Parity
  691/691** (gate ON and OFF). **Revised forward direction (per the user):** the WRITE side is
  ONE `register_type(info)` fn that branches on the kind IN THE COMPILER (subsuming `define`'s
  per-kind dispatch), not a per-kind `register_struct`.
- **Phase 3 P3.2 (VM plan) — field-level reflection readers: `type_nominal_name` + `type_field_name` + `type_field_type` (2026-06-18).**
  Three more `compiler`-library readers on the same `TypeId`-handle shape (added to
  `compiler_lib.bound_fns` AND `Vm.callCompilerFn`), each backed by a new `TypeTable` query
  BOTH paths call (no drift): `nominalName` (a named type's own name handle; loud-bail for
  unnamed types like `i64`/pointers), `memberName` (struct/union/tagged-union field, enum
  variant, named-tuple element), `memberType` (struct/tuple/array/vector member type). All
  loud-bail on out-of-range idx / no-member (no silent default). First MULTI-ARG compiler
  fns — `callCompilerFn` reads arg 1 = idx; added `Vm.argHandle`/`argTypeId` (range-checked
  u32/TypeId arg reads) and refactored `find_type`/`type_field_count` onto them. Named
  `type_*` to avoid clashing with the std metatype builtins (`field_name`/`type_name` exist
  in core.sx); `nominalName` (the TypeTable method) is distinct from the existing
  `typeName(id) []const u8` display-string renderer. Example `0629-comptime-compiler-field-reflect`
  reflects `Pair { lo: Point; hi: Point }` — each field name + the nominal name of a field's
  type, all `#run`-folded, all VM-HANDLED natively. VM unit test added (type_field_name → "hi";
  type_nominal_name(type_field_type(Pair,0)) → "Point"). **Parity 690/690** (gate ON and OFF).
- **Phase 3 P3.1 (VM plan) — first read-only reflection readers: `find_type` + `type_field_count` (2026-06-18).**
  Two more `compiler`-library fns, bound the same way as the `intern`/`text_of` seed (added
  to `compiler_lib.bound_fns` for the legacy handler + the welded-decl export check, AND to
  `Vm.callCompilerFn` for the native flat-memory path — NO marshaling). A **type handle is a
  plain `u32` `TypeId`** (like `StringId`), so both keep the seed's clean scalar shape:
  `find_type(name: StringId) -> TypeId` (`TypeTable.findByName`, `unresolved`/0 if absent) and
  `type_field_count(t: TypeId) -> i64` (a NEW `TypeTable.memberCount` query — struct/union/
  tagged-union fields, enum variants, array/vector length — called by BOTH paths so they
  can't drift; bails loudly, never a silent 0). New example `0628-comptime-compiler-find-type`
  chains `intern → find_type → type_field_count` (and a not-found lookup → 0), both folded at
  `#run`, both VM-HANDLED natively (trace confirms no fallback). VM unit test added
  (`find_type` + `type_field_count`, struct found → 3 fields, missing → `unresolved`).
  **Parity 689/689** (gate ON and OFF). **Decision (resolves the plan's `find_type → ?Type`
  sketch):** return a NON-optional `TypeId` with the `unresolved` (0) sentinel for not-found,
  NOT `?Type` — a `Type` value resolves to `.any` (which the flat-memory VM doesn't represent)
  and an optional can't cross the legacy↔VM eval boundary; `unresolved` is the project-blessed
  unmistakable "no type" marker. Forward (P3.2): more readers on the same handle shape
  (`type_name`/`field_name`/`field_type`/kind), then `register_struct` (first mutating fn).
- **VM robustness — `Frame` bounds-check; lowering-time `#insert` wiring explored + reverted (2026-06-18).**
  Explored wiring the VM at the LOWERING-time comptime site (`evalComptimeString`, the
  `#insert` string fold). 12/13 `#insert` examples ran on the VM with parity, but `0737`
  (an `#insert` of an unresolved `secret()`) CRASHED the VM (SIGABRT): lowering-time IR can
  be malformed (a `ret Ref.none` from the unresolved name) and `Frame.get` panicked on the
  out-of-range index. **Decision: reverted the lowering-time wiring** — unlike the emit-time
  folds (fully lowered IR), lowering-time IR can be erroneous, and hardening the VM against
  ALL malformed IR (every `ref_types[...]` / `aggType` access, not just `Frame`) is out of
  scope here. The emit-time sites already give full corpus coverage. **KEPT** the defensive
  fix regardless (CLAUDE.md "never crash"): `Frame.get`/`set` now bounds-check and flip a
  `bad_ref` flag; the `run` loop bails (`badRef`) instead of panicking. Unit test added
  (malformed `ret Ref.none` → bail, not crash). Parity **688/688** both ways.
- **Phase 3 SEED (VM plan) — compiler-call path: `intern`/`text_of` native on the VM (2026-06-18).**
  `invoke` now dispatches a welded `compiler`-library fn (gated on `compiler_welded`) to
  `Vm.callCompilerFn`, serviced NATIVELY on flat memory (no legacy `Interpreter`):
  `intern(string)->StringId` reads the flat-memory string bytes and `internString`s into the
  const-cast table (pool-only — doesn't touch type layout, so cached sizes stay valid);
  `text_of(StringId)->string` materializes the pooled text back into flat memory. Unlocked
  `0626`; the ONLY remaining const-init fallback is now the inline-asm global (`1654`).
  Parity **688/688** (gate ON and OFF); unit test added. This is the mechanism Phase 3 grows
  — the next compiler fns (`find_type`, `register_struct`, reflection readers) bind the same
  way (flat-memory pointer in, handle/pointer out, no marshaling).
- **Phase 1.final step 9 (VM plan) — `-Dcomptime-flat` build flag (the "swap behind a build flag" step) (2026-06-18).**
  Added the `-Dcomptime-flat` build option (build.zig → a `build_opts` options module on
  `mod`; `emit_llvm.init` reads `build_opts.comptime_flat or SX_COMPTIME_FLAT env`). This is
  the plan's "reach parity → swap behind a build flag → delete the old path" mechanism.
  `zig build test -Dcomptime-flat` runs the FULL corpus on the VM (688/0). Verified the flag
  toggles the binary: flag-built `sx` reports VM HANDLED with no env var; default-built does
  not. Default OFF — `zig build test` unchanged (688/0). Env var still works for ad-hoc runs.
  Next (forward): Phase 2 (bytecode) / Phase 3 (compiler-API on flat memory); eventual
  default-flip + legacy deletion.
- **Phase 1.final step 8 (VM plan) — wire the `#run` side-effect path + trace-clear-on-fallback (2026-06-18).**
  Wired the SECOND comptime call site (`runComptimeSideEffects`, top-level `#run <expr>;`)
  through `tryEval` with legacy fallback, mirroring the const-init fold. `tryEval` now
  handles void/noreturn entries (→ `.void_val`) so a void side-effect doesn't bail at the
  result conversion. **Fixed a trace-corruption** the new site exposed (`1035`): a
  side-effect that pushes return-trace frames and then bails (e.g. on `print`) had the
  legacy re-run DOUBLE-push them (`sx_trace_push` is a side effect on the shared buffer).
  Both wiring sites now `sx_trace_clear()` right before the legacy fallback, discarding the
  VM's partial pushes. **Parity 688/688** (gate ON and OFF). Most side-effects still bail
  (print/global_addr/call_builtin) → legacy, but the path is now uniform. All comptime
  evaluation routes through the VM-with-fallback.
- **Phase 1.final step 7 (VM plan) — is_comptime + failable/error cluster + signed-load fix; coverage 31→36 (2026-06-18).**
  `is_comptime` → 1 (unlocked `1030`). Ported the failable/error-channel cluster (`1037`
  escape, `1038` handled): `kindOf(error_set)→word`, `regToValue` bridges TUPLES (the
  failable `(value…,tag)` shape `checkComptimeFailable` reads), `trace_frame` packs
  `(func_id<<32|span.start)` from a new `call_stack` (pushed by invoke/runEntry), and
  `sx_trace_push`/`sx_trace_clear` serviced NATIVELY (the VM calls the real sx_trace.c
  functions linked into the compiler, so the return-trace buffer is populated identically
  to legacy). raise/catch/or now run on the VM. **Surfaced + fixed a real GENERAL bug:**
  `readField` was ZERO-extending signed sub-64-bit loads, so a stored `i32 -1` reloaded as
  `0xFFFFFFFF` (+4.29e9) and `< 0` was false — silently hiding `raise error.Bad`; now
  SIGN-extends `i8`/`i16`/`i32`/`isize` (gate-ON parity confirms it's a strict fix; unit
  test added). VM HANDLES **36** corpus const-inits (was 31); **parity 688/688** (gate ON
  and OFF). Only **2 fallbacks** remain, both principled: `intern` (`0626`, welded
  compiler-API fn — Phase 3) + inline-asm global (`1654`). Forward work: Phase 2 (bytecode),
  Phase 3 (compiler-API on flat memory).
- **Phase 1.final step 6 (VM plan) — real default context + call_indirect + func_ref + global_get; coverage 27→31 (2026-06-17).**
  Per the user's direction ("the VM can set up a default context"), `runEntry` now
  materializes the REAL default context instead of a zeroed one. The implicit-ctx param is
  an opaque `*void`, so `materializeDefaultContext` finds the `__sx_default_context` global
  and lays its initializer (`{ {null, alloc_fn, dealloc_fn}, null }`, the CAllocator thunk
  func-refs) into flat memory via a new recursive `layoutConst`. With `func_ref` (function
  value encoded `FuncId.index()+1`, reserving word 0 for the null fn-ptr) and
  `call_indirect` (decode word → FuncId → dispatch; 0 → bail) ported, the whole allocator
  protocol runs on the VM:
  `context.allocator.alloc_bytes` → call_indirect → thunk → `CAllocator.alloc_bytes` →
  `libc_malloc` → native flat malloc. Unlocked `0606` (string global). Also: `global_get`
  lazily evaluates a comptime global's `comptime_func` (memoized) — unlocked `CT_CHAIN`;
  field access (`fieldOffset`/`struct_get`) handles string/slice `{ptr@0,len@8}` fat
  pointers (needed by `alloc_string`); `regToValue` maps function-typed words → `.func_ref`
  (kept `1128`'s rejection byte-identical). Native `malloc` is still required (the thunk
  bottoms out at it; a host pointer can't be used with flat-memory load/store). VM HANDLES
  **31** corpus const-inits (was 27); **parity 688/688** (gate ON and OFF). Unit tests:
  global_get, func_ref+call_indirect. Remaining fallbacks (7): `.unsupported` aggregates
  (3× — `1037`/`1038`), extern/builtin `intern`+asm (2×), `trace_frame`, `is_comptime`.
- **Phase 1.final step 5 cont. (VM plan) — libc memory builtins + f32 fix; coverage 16→27 (2026-06-17).**
  Identified the dominant fallback (`call to extern/builtin`) as **11× `malloc`** (0604) +
  1× `intern`. Modeled a curated set of libc MEMORY builtins natively on flat memory
  (`Vm.callMemBuiltin`): `malloc`/`calloc` → `allocBytes` (16-aligned, 256-MiB cap → bail),
  `free` → no-op, `memcpy`/`memmove`/`memset` on flat bytes — sandboxed (no host heap/dlsym),
  target-aware; the computed result is byte-identical to legacy (which calls real libc).
  This surfaced a **real latent f32 bug**: float registers hold f64 bits, but f32 MEMORY is
  the 4-byte single — `readField`/`writeField` were truncating the f64 bits (writing zeros
  for `1.0`); now they `@floatCast` on f32 load/store (mirrors legacy `storeAtRawPtr`).
  Result: VM HANDLES **27** corpus const-inits (was 16); **parity 688/688** (gate ON and
  OFF). Unit tests added (f32 round-trip; malloc → usable flat memory). Next: the `kindOf`
  `.unsupported` aggregates (3×), `global_get` (2×), the rest.
- **Phase 1.final step 5 (VM plan) — implicit-context materialization; coverage 0→16 (2026-06-17).**
  `tryEval` now MATERIALIZES the implicit ctx instead of skipping it: a `has_implicit_ctx`
  comptime entry (sole param `*Context`) gets a zeroed `Context` of the right size/align
  in flat memory, its address passed as arg 0. Const bodies that ignore the ctx run; a
  body that uses the allocator hits unported `call_indirect` → bails → legacy. No func-ref
  materialization needed (handled bodies don't read ctx contents; parity is the guard).
  Fixed a real bug surfaced by the coverage pass: storing a `null` non-pointer optional
  (the `null_addr` sentinel) into an aggregate slot OOB-bailed — `writeField` now ZEROES
  the destination for a `null_addr` aggregate source (= none/empty); unit-test regression
  added. Result: VM HANDLES **16** corpus const-inits (was 0); **parity 688/688 both
  gate ON and OFF**. Next: port the ops the trace names — `call_builtin`/`compiler_call`/
  extern (13×, via the bridge), `kindOf` `.unsupported` aggregates (3×), `global_get` (2×),
  func_ref / call_indirect / trace_frame / is_comptime.
- **Phase 1.final steps 1–4 (VM plan) — host wiring landed; coverage measured (2026-06-17).**
  (1) **Hardening:** `Machine.readWord`/`writeWord`/`bytes` now return `error.OutOfBounds`
  (null / out-of-range / oversized / overflow-safe) instead of `assert`-panicking;
  `OutOfBounds` added to `Vm.Error`; `try` threaded through every helper + exec arm + the
  bridge. New unit tests (accessor OOB returns; null-deref → `tryEval` null, not a crash).
  (2) **Implicit context:** `tryEval` returns null for `has_implicit_ctx` funcs (legacy
  fallback) — conservative; full ctx materialization deferred to step 5. (3) **Wiring:**
  const-init fold in `emit_llvm.zig` `emitGlobals` is `(if comptime_flat) tryEval else
  null) orelse interp.call(...)`, gated by env `SX_COMPTIME_FLAT` (read once into
  `LLVMEmitter.comptime_flat`). Default OFF. (4) **Parity + coverage:** gate ON → full
  corpus byte-identical (688, 0 failed) + manual 0605/0606/0607 byte-identical.
  **Finding: 0 of 37 measured corpus const-inits are VM-handled — ALL are
  `has_implicit_ctx`-gated.** Added a coverage-trace facility (`comptime_vm.last_bail_reason`
  + env `SX_COMPTIME_FLAT_TRACE`). **Next: step 5 = implicit-context materialization** (the
  unblocker), then port the deferred ops. 688 corpus green (gate OFF).
- **Phase 1.final start (VM plan) — wiring entry point `tryEval` (2026-06-17).**
  `comptime_vm.tryEval(gpa, module, func_id) ?Value` runs a comptime function entirely on
  the VM, returns a legacy `Value` (deep-copied to `gpa`) or `null` to fall back.
  Unit-tested (pure 6*7 → 42; unbox_any → null). NOT yet routed into the host: needs
  (1) panic→error hardening of `Machine` accessors so arbitrary funcs bail instead of
  crashing, (2) implicit-ctx handling, (3) wiring at `emit_llvm` const-init behind
  `SX_COMPTIME_FLAT`, (4) corpus parity run. See `PLAN-COMPILER-VM.md` Phase 1.final.
  688 corpus green.
- **Phase 1 sub-step 1.5b (VM plan) — Reg↔Value boundary bridge (2026-06-17).**
  Builtin/compiler_call/extern handlers are coupled to the legacy `Interpreter`, so the
  wiring will use WHOLE-FUNCTION fallback (VM runs pure functions; bail → legacy re-runs
  the whole eval). Built the boundary bridge that enables it: `valueToReg` (Value arg →
  Reg, aggregates into flat memory) + `regToValue` (VM result → Value, deep-copied).
  Covers scalars/strings/structs; other shapes bail. Transitional. Round-trip
  unit-tested. 688 corpus green. Next: the wiring (flag + route a comptime entry through
  the VM with legacy fallback).
- **Phase 1 sub-step 1.5 (VM plan) — direct `call` + stack-lifetime change (2026-06-17).**
  `Vm` gained `module` (callee resolution) + `depth`/`max_depth` guard. `call` marshals
  arg Refs → Reg and recursively runs the callee; aggregates pass as Addrs over shared
  flat memory. `Frame` no longer reclaims the machine on exit (else a returned aggregate
  Addr dangles) — allocations live to `Vm.deinit`. Extern/builtin callees bail (1.5b).
  Unit-tested: direct call (142), recursion sum(0..n) (15/55). 688 corpus green. Next:
  1.5b (call_builtin/compiler_call/extern), then hybrid wiring.
- **Phase 1 sub-step 4d (VM plan) — deref/addr_of; pivot decision (2026-06-17).**
  Ported `addr_of` (pass-through) + `deref` (readField through pointer), unit-tested
  (deref *i64 → 77, addr_of struct + field → 80). DECIDED to stop porting rarer ops
  (tagged-union payload/any/closures) blind — their byte semantics are ambiguous without
  real call sites — and pivot to CALLS (sub-step 1.5: `call`, then builtin/compiler) +
  HYBRID WIRING (`-Dcomptime-flat` → VM with legacy fallback on `error.Unsupported`), so
  the VM runs the real corpus and surfaces exactly what's needed. Key design point for
  calls: aggregate-return lifetime → drop per-frame stack reclaim (let a comptime eval's
  allocations live to `Vm.deinit`). 688 corpus green. See `PLAN-COMPILER-VM.md` decision
  block.
- **Phase 1 sub-step 4c (VM plan) — optionals + payloadless enums (2026-06-17).**
  `kindOf`: enum → word; `?T` → word (pointer-child, null==0) or `{T@0,i1@sizeof(T)}`
  aggregate. Ported optional_wrap/unwrap/has_value/coalesce (`optChildIsPtr`/`optHas`;
  const_null reads as none) + payloadless enum_init/enum_tag. Unit-tested (?i64 → 91,
  ?*i64 null==0 → 99, enum tag → 11). 688 corpus green. Next: 4d (tagged unions, any,
  closures).
- **Phase 1 sub-step 4b (VM plan) — slices + strings on flat memory (2026-06-17).**
  `{ptr@0(pointer_size), len@8(i64)}` fat pointers (kindOf: string/slice → aggregate).
  Ported `const_string` (text+NUL + fat pointer in flat memory), `length`/`data_ptr`,
  `array_to_slice`, `subslice`, index-through-slice (`elemAddr` loads `.ptr`), and
  `str_eq`/`str_ne` (memcmp). Unit-tested (str length+eq/ne, array→slice index sum=23,
  subslice sum=43). 688 corpus green. Next: 4c (optionals/enums/any/closures).
- **Phase 1 sub-step 4a (VM plan) — tuples + arrays on flat memory (2026-06-17).**
  `kindOf` widened (tuple/array → aggregate). Ported `tuple_init`/`tuple_get`
  (`tupleFieldOffset`), `index_get`/`index_gep` (`elemAddr` = base + idx*elem_size over
  array/pointer/many_pointer; slice/string bases bail), `length` on array values.
  Unit-tested (mixed tuple, [3]i64 index sum=42, length=3). 688 corpus green. Next:
  sub-step 4b (slices/strings, then optionals/enums/any/closures).
- **Phase 1 sub-step 3 (VM plan) — memory + structs on flat memory (2026-06-17).**
  `Vm` gained optional `table: *const TypeTable` (target-aware layout). Ported
  `alloca`/`load`/`store` + `struct_init`/`struct_get`/`struct_gep`, laying structs out
  at the table's natural offsets. Value model: scalar/pointer → register word;
  struct → lives in flat memory, its value IS its address (read→addr, write→memcpy), so
  nested structs compose and `struct_gep` = base+offset. `kindOf` bails loudly on
  not-yet-ported types. Addr-based values survive allocator realloc. Unit-tested
  (struct round-trip, alloca+gep+store+load, nested struct). 688 corpus green. Next:
  sub-step 4 (arrays/slices/strings/optionals/enums/tuples/any/closures, then calls).
- **Phase 1 sub-step 2 (VM plan) — flat-memory executor: scalars + control flow
  (2026-06-17).** Added `Vm` to `comptime_vm.zig`: walks the same IR `Inst` over
  flat-memory frames (register `Reg` = scalar bits or `Addr`), mirroring the legacy
  interp's scalar semantics (i64 wrapping/signed, f64). Ported constants, arithmetic,
  comparison, logical, conversions, terminators (`br`/`cond_br`/`ret`/`ret_void`) and
  `block_param`; every other op bails loudly (`error.Unsupported` + op name in
  `detail`). Unit-tested on hand-built tiny IR (`Fb` builder): int add, f64 arithmetic,
  cond_br selection, a block-param loop, div-by-zero + unsupported-op bails. Corpus
  untouched (688 green). Next: sub-step 3 (memory + aggregates on flat memory, where
  target-aware layout enters).
- **Phase 1 sub-step 1 (VM plan) — flat-memory machine substrate (2026-06-17).**
  New `src/ir/comptime_vm.zig`: `Machine` (linear byte memory + bump/stack allocator
  with `mark`/`reset`, scalar `readWord`/`writeWord` 1/2/4/8 LE, `bytes` views, addr 0
  reserved as `null_addr`) + `Frame` (Ref-indexed register file, stack reclamation on
  deinit). `Reg` = raw u64 (immediate scalar OR `Addr`). Unit-tested
  (`comptime_vm.test.zig`), registered in the barrel; standalone — the legacy
  interpreter stays live, corpus untouched (688 green). Next: sub-step 2 (executor +
  scalar/branch ops over the same IR). Also removed the "~500 lines / split step" rule
  from CLAUDE.md per request.
- **Phase 0 (VM plan) — struct-weld stripped; `intern`/`text_of` bridge kept
  (2026-06-17).** Removed the struct-weld registry from `compiler_lib.zig`
  (`weldStruct`/`bound_types`/`BoundType`/`FieldLayout`/`findType`/`SxField`/
  `LayoutMismatch`/`validateStructLayout`), `validateWeldedStruct`/`weldedFieldOrderStr`
  + the `sd.abi == .zig` call from `nominal.zig`, the struct-weld unit tests, and
  examples `0625`/`0627`/`1183`/`1186`. KEPT (decision) the `intern`/`text_of` function
  host-call bridge — a clean scalar dispatch, not weld/serialize/marshal, the Phase-3
  compiler-call seed — so `weldedCompilerFn`, the `compiler_welded` dispatch, the
  `emitCall` comptime-only gate, the `#library`/`abi`/`extern` syntax, and examples
  `0626`/`1184`/`1185` remain. `zig build test` green (688 corpus, 0 failed). Next:
  Phase 1 (flat-memory value model) per `PLAN-COMPILER-VM.md`.
- **DIRECTION CHANGE — pivot off the byte-weld to a flat-memory bytecode VM
  (2026-06-17).** Decided the weld + serialization/marshaling bridge is the wrong
  direction (it hand-marshals onto a comptime value model that isn't bytes — exactly
  what the design set out to kill). New foundation: a bytecode VM over flat memory so
  comptime values are native bytes; the compiler-API then rides on it via direct memory
  (no weld/validation/marshaling). JIT-native comptime was weighed and rejected (breaks
  cross-compilation, loses the sandbox). Wrote `current/PLAN-COMPILER-VM.md` (Phase 0
  strip → Phase 1 flat-memory value model → Phase 2 bytecode → Phase 3 compiler-API on
  flat memory). Banner added to `design/comptime-compiler-api.md` (superseded). Reverted
  the session's uncommitted `register_struct`/`find_type` marshaling experiment back to
  `reify` HEAD (40d075c). No code stripped yet — Phase 0 is the next action.
- **Phase 2 — welded structs by reflection + memory-order validation.** Dropped
  the byte-layout-override engine (computeWeldPlan / offset-ordered LLVM struct /
  byte-blob — all explored, all unnecessary). Instead: the sx header declares
  fields in the compiler type's memory order; the compiler reflects the bound Zig
  type (`@typeInfo`/`@offsetOf`/`@sizeOf`) and validates the header matches with
  loud diagnostics (field-not-found, wrong-order+expected-order, size mismatch).
  On pass it's an ordinary byte-identical struct — cast + deref just works.
  Examples 0627 (usable) / 1186 (wrong-order diagnostic). Suite green (692).
- **Phase 2.1 — weld-plan layout math (REMOVED).** The byte-layout-override math;
  superseded by the reflection+validation design and deleted.
- **Phase 1 polish — comptime-only enforcement.** A runtime call to a welded fn is
  a clean build-gating error (`emitCall` gate, guarded by enclosing-`is_comptime`
  so `#run`/`::` uses stay green), not a link failure. Example 1185. Build + suite
  green (458 unit, 690 corpus).
- **Phase 1.1 fifth sub-step — host-call bridge (welded functions).**
  `compiler_lib` function registry (`intern`/`text_of`) + `findFn`; IR `Function`
  `compiler_welded` flag set/validated in `declareFunction` (`weldedCompilerFn`);
  `interp.call()` dispatches welded calls to the Zig handler. Examples 0626 (round-
  trip) + 1184 (unexported-fn diagnostic); `findFn` unit-tested. Runtime-call clean
  rejection deferred (loud link error today). Build + suite green (458 unit, 689
  corpus).
- **Phase 1.1 fourth sub-step — welded-struct layout validation.**
  `validateStructLayout` (pure, unit-tested) + `validateWeldedStruct` wired into
  `registerStructDecl`: a `struct abi(.zig) extern compiler` is validated against
  the registry (lib == compiler, name exported, layout matches) with build-gating
  diagnostics. `#library "compiler"` no longer dlopen'd. Examples 0625 (faithful
  Field) + 1183 (field-count mismatch diagnostic). Offset-override/GEP deferred to
  Phase 2 (not exercised by Field's natural layout). Build + suite green (456 unit,
  687 corpus).
- **Phase 1.1 third sub-step — binding registry.** New `src/ir/compiler_lib.zig`:
  the `compiler` lib's welded-type registry; `Field` welded to
  `StructInfo.Field` with layout baked from the real Zig type
  (`@offsetOf`/`@sizeOf`/`@alignOf`); `findType` lookup proven by unit test
  (+ null off the export list). Standalone island — not yet consumed by lowering.
  Build + suite green (454 unit tests). Break-verified.
- **Phase 1.1 second sub-step — struct-decl binding parses.** `ast.StructDecl`
  gained `abi` + `extern_lib`; `parseStructDecl` parses `abi(.zig) extern <lib>`
  after `struct`. Parser unit tests (welded `Field` + plain struct), break-verified.
  Build + suite green. Parse-only sub-step (fns + structs) of Phase 1.1 complete.
- **Phase 1.1 first sub-step + `callconv`→`abi` unification.** Parsed `abi(.zig)
  extern <lib>` on fn decls; unified `callconv` into `abi(.c|.zig|.pure)` (removed
  the `callconv` keyword), migrated 52 sx files + compiler diagnostics + docs +
  snapshots. Build + suite green. The original design's `extern(.zig)` single
  qualifier was split into `abi(.zig)` (ABI/layout, before extern) + `extern
  <lib>` (linkage + source) — recorded in the design doc's syntax-decision note.