sx/current/CHECKPOINT-MEM.md

# Memory Module — Progress + Issues Log

Tracking checkpoint for the mem.sx Zig-aligned implementation
(plan: `~/.claude/plans/tidy-doodling-cray.md`).

## Last completed step

- **`resolveType(null) → .s64` silent fallback removed.** `resolveType`
  now takes a non-optional `*const Node`; the `null → .s64` branch is
  gone. Callers that legitimately had no annotation handle it
  themselves: top-level `var_decl` at `lower.zig:630` infers from the
  initializer or diagnoses if neither is present (matches the
  `lowerVarDecl` pattern that already existed for locals).
  `FfiIntrinsicCall.return_type` is non-optional in the AST so its
  callers (`#objc_call`, `#jni_call`) didn't actually need the
  fallback. JNI super-call and JNI method paths were already guarded
  with `if (rt) |t| ... else .void`.

  Caller cleanup also dropped the explicit `if (x != null)` guards in
  favour of optional-payload syntax (`if (cd.type_annotation) |ta|
  ...`), which makes the always-non-null path obvious from the type.

  Real-world impact: `g_pi := 3.14;` at the top level used to be
  silently typed as `s64`. Now it infers as `f64`. Regression at
  `examples/137-toplevel-var-type-inference.sx` (count/pi/flag — int /
  float / bool inferred correctly). 159/159 example tests + chess
  clean.

- **Phase 1.4a — IR `TypeId` threaded through `valueToLLVMConst`;
  string/slice fat-pointer aggregates serialize by reading host
  memory.** The Phase 1.4 serializer bailed on `heap_ptr` / `byte_ptr`
  and silently emitted `i0 0` for the trap-case where a `.int` host
  address landed in a ptr-typed slot. Now the call site at
  `emit_llvm.zig:676` passes `global.ty` (TypeId) and `&interp_inst`
  instead of just the LLVM type. The serializer splits on the IR
  type:

  - `string` / `slice` (fat pointer `{data, len}`): extract `len`,
    read that many bytes from the data field's address (heap_ptr →
    `interp.heapSlice`; byte_ptr/int → raw process memory via a new
    `readHostBytes` helper; string literal → direct slice). Emit
    the bytes as a private global byte array via the existing
    `emitConstStringGlobal` and use it as the aggregate's data ptr.
  - `struct`: walk the IR field types in lockstep with the value
    fields; recurse per field with its declared TypeId. Replaces
    the old LLVM-type-walk via `LLVMStructGetTypeAtIndex` which
    couldn't tell `string`-typed fields from generic ptr fields.
  - `array`: walk elements with the element TypeId.

  The `.int → ptr` slot mismatch (a host address landing in a
  non-fat-pointer ptr slot) now bails loudly with a named
  diagnostic — that's the genuine heap-walk frontier where future
  work would need to capture struct content recursively, not the
  silent malformed-const we had before.

  `Interpreter.heapSlice` was promoted from package-private to
  `pub` so the serializer can read interp heap. Regression at
  `examples/136-comptime-string-global.sx`: `GREETING :: #run
  build_greeting();` where `build_greeting` returns `concat("hello",
  " world")` — runtime prints `greeting = 'hello world' / greeting.len
  = 11`. Pre-1.4a this segfaulted. 158/158 example tests + chess
  clean on all three platforms via `tools/verify-step.sh`.

- **Allocator `init` returns the state by value.** Building on the
  Option 3 lvalue-borrow rule, `GPA.init`, `Arena.init`, and
  `TrackingAllocator.init` now return `T` (not `*T`). The caller binds
  the local; the local IS the storage. `xx local` borrows under Option
  3 so the `Allocator` protocol value's `ctx` points at the local.
  Saves one `libc_malloc` per allocator instance; closes the
  state-struct leak surface (the local goes away with its scope, no
  explicit `deinit` needed for the struct itself).

  Migration:
  - `library/modules/allocators.sx`: `init` signatures changed
    (`-> T` instead of `-> *T`); `Arena.deinit` drops the trailing
    `parent.dealloc(xx a)` — caller owns the storage.
  - `library/modules/ui/pipeline.sx`: `arena_a: Arena;` /
    `arena_b: Arena;` (was `*Arena;`); `@self.arena_a` /
    `@self.arena_b` at the use site that needs `*Arena` for the
    `build_arena` local.
  - `examples/126-xx-recover-then-dispatch.sx` updated: comparisons
    against `*GPA` use `@gpa` now.
  - `examples/135-xx-lvalue-borrows.sx` simplified: no
    `tracker_ptr.*` deref needed.
  - `examples/50-smoke.sx` Arena counts dropped by 1 each (no
    state-struct alloc); comments + snapshots updated to match.
  - CLAUDE.md "Allocator construction" rule rewritten around the
    by-value convention.

  157/157 example tests + chess clean on macOS / iOS sim / Android
  (`tools/verify-step.sh` ran green).

- **`xx <lvalue>` borrows the operand's storage** (Option 3 in the
  protocol-erasure design discussion). Today's behavior — `xx
  <struct-typed local>` heap-copies the value — was a silent footgun:
  the protocol value pointed at the heap copy, the original local
  stayed stale, mutations through the protocol weren't visible to the
  original (and vice versa). Under the new rule, when the operand
  names existing storage (identifier, field access, index expression,
  dereferenced pointer), `xx` takes its address and the protocol
  borrows. Heap-copy is reserved for `xx <rvalue>` — struct literals,
  function-call results, arithmetic expressions, anything without its
  own storage.

  Single point of change at `buildProtocolErasure` in `lower.zig:10334`,
  via a new `isLvalueExpr` helper at `lower.zig:10322`. specs.md §3
  ownership table updated. The `examples/130-...` regression that
  previously tested heap-copy on `xx <local>` now tests `xx
  <struct-literal>` (still the heap-copy path); new regression
  `examples/135-xx-lvalue-borrows.sx` witnesses the borrow path via
  TrackingAllocator. 157/157 example tests + chess clean across all
  three platforms (`tools/verify-step.sh` gate ran green right
  before this work landed).

- **Phase 1.4 — `valueToLLVMConst` upgraded to handle every interp
  `Value` variant.** The serializer at `emit_llvm.zig:734` used to
  collapse anything past int/float/boolean into `LLVMConstNull(ty)` —
  a silent fallback that emitted `{0, 0}` for any `#run` returning a
  struct, string, function pointer, or pointer. Replaced with a real
  walk: int/float/boolean as before; null_val → LLVMConstNull;
  void_val/undef → LLVMGetUndef; func_ref → func_map lookup;
  string → emitConstStringGlobal; aggregate → recurse via
  `LLVMStructGetTypeAtIndex` / `LLVMGetElementType`. The unsupported
  cases (heap_ptr/byte_ptr/slot_ptr/closure/type_tag) now bail loudly
  with a named diagnostic that includes the global name (per CLAUDE.md
  REJECTED PATTERNS — no silent unimplemented arms). Heap_ptr is
  deferred to Phase 1.4a (needs IR TypeId threaded down for cyclic /
  recursive heap content).

  Also closes the type-inference half of the same bug: `NAME :: #run
  expr;` with no annotation used to default to `s64` (silent fallback
  in `resolveType(null)`). `lowerComptimeGlobal` now infers from the
  expression's return type when no annotation is provided. The
  silent fallback in `resolveType` itself is left in place for other
  callers — separate audit, separate session.

  Regression test at
  `examples/134-comptime-aggregate-global.sx`: a `POINT :: #run
  make_point();` struct binding now prints the real fields instead
  of zeros, on both interp and codegen paths. 156/156 example tests
  + chess clean.

## Current state

Phase 0.0c shipped (allocator API on one-line `init` returning `*T`;
TrackingAllocator added). 148/148 tests pass.

Phase 0.2 verified across all known patterns. Phase 0.3 documented.

Phase 0 of the MEM plan is **COMPLETE**.

Fixed this session by paired sessions: issue-0038 (transitive
`#import`), issue-0039 (chess + stdlib migration to explicit imports),
issue-0040 (generic struct method dot-dispatch). All landed and
re-verified. Full gate green: 151/151 example tests + chess on macOS
/ iOS sim / Android with screenshots.

Phase 0 spike outcomes:
- **0.2** xx cast patterns — verified across 5 additional shapes
  beyond the issue-0037 regression.
- **0.3** chess allocator audit — documented; no chess-side
  migration needed.
- **0.6** `align_of($T)` builtin — landed. Mirrors `size_of` in
  inst.zig BuiltinId, lower.zig (registry + return-type + reflection
  handler), interp.zig (fallback), sema.zig (allowed-builtins list),
  lsp/server.zig (both completion tables), library/modules/std.sx.
  Smoke coverage added in `examples/50-smoke.sx` (u8/s32/s64/Point).
- **0.7** `#import` transitivity — surfaced and fixed via issue-0038.
- **0.8** `#foreign("c")` rename syntax — confirmed
  `#foreign libc "name"`.
- **0.9** generic UFCS dispatch — surfaced and fixed via issue-0040.
  Dot-call now works for struct methods with `$T: Type` parameters,
  so the plan's `a.create(MyType)` shape is viable.

issue-0041 (pointer type as type-arg) and issue-0042 (alias names in
`resolveTypeArg`) **FIXED**. Regression tests at
`examples/issue-0041.sx` and `examples/issue-0042.sx`. Scratch
verification: `size_of(*u8)=8`, `size_of(Ptr where Ptr::*u8)=8`,
`size_of(?u8)=2`, `size_of(Maybe where Maybe::?u8)=2` — all clean on
interp + codegen.

Also landed during 0041/0042: the silent `.s64` fallback in
`resolveTypeArg` is gone — unresolved type names now emit a real
diagnostic. Surfaced and removed two bogus `size_of(Complex)` /
`size_of(Sx)` calls in `examples/10-generic-struct.sx` that were
relying on the silent default. Caller-side speculative paths in
`buildTypeBindings` + `inferGenericReturnType` now gate the call
with `type_bridge.isTypeShapedAstNode` before invoking
`resolveTypeArg`. CLAUDE.md REJECTED PATTERNS gained a section
forbidding silent `orelse default` returns in compiler code.

Parser regressions introduced by the 0041 work fixed in
`src/parser.zig:hasFnBodyAfterArrow`:
- `(s: string) -> [:0]u8 { ... }` — added `.colon` to the
  return-type token walk.
- `(x) -> Closure(...) -> R { ... }` — added `.arrow` so nested
  return-type arrows continue the walk.
- `name :: (self: T) -> Ret;` inside `struct #compiler` — recognise
  trailing `;` as a method-decl terminator. Was silently dropping
  every `BuildOptions.*` method from `fn_ast_map`.

Full gate green: 151/151 example tests + chess on macOS / iOS sim /
Android with screenshots.

**Implicit Context refactor SHIPPED.** Every default-conv sx function
now carries `__sx_ctx: *void` at LLVM slot 0. `context.X` lookups
resolve through the lowering's `current_ctx_ref` — a one-indirection
load, no per-access walk, thread-safe by construction (each call
stack carries its own Context chain). `push Context.{...}` allocates
a fresh slot and rebinds `current_ctx_ref` for the body's lexical
scope. The `context` LLVM global is gone; the only runtime Context
is the static `__sx_default_context` (a CAllocator backing libc
malloc/free), installed at FFI-inbound entries (`main`, `Java_*`,
`JNI_OnLoad`) and used by the interp for `#run` evaluation.

What landed:
- `Function.has_implicit_ctx` flag + `Module.has_implicit_ctx`
  per-compilation switch (gated by `Context :: struct {...}` being
  present in the dep graph).
- Param prepend + call-site forwarding across every sx-to-sx path:
  direct calls, indirect through fn-pointer vars, protocol dispatch,
  closure trampolines, lambda trampolines, bare-fn trampolines,
  generic monomorphizations, comptime functions.
- `ConstantValue.func_ref: FuncId` variant so the static initializer
  for `__sx_default_context` can reference the CAllocator thunks.
- emit_llvm two-pass global emission: aggregates that name funcs by
  FuncId resolve them after `func_map` is populated.
- Interp: `defaultContextValue()` builds the Context aggregate on
  demand; `interp.call` bootstraps slot_ptr(0) when an entry function
  with implicit ctx is invoked sans args; `materializeCtxArg` derefs
  the caller's slot_ptr at every sx-to-sx boundary so callees can
  treat ref 0 as the Context aggregate; `.load` of an aggregate is
  passthrough; `.global_addr` of `__sx_default_context` returns the
  aggregate directly.
- `matchContextAllocCall` is GONE (commit `d415bcc`). Comptime now
  runs the full Allocator-protocol dispatch chain — the same IR
  codegen emits — by reusing the parent module instead of spinning
  up a separate `ct_module`. The interp gained raw-pointer paths
  (`index_gep`, `index_get`, `store`, `marshalForeignArg`,
  `asString`) so `context.allocator.alloc` bottoms out at host
  `libc_malloc` and the returned pointer survives downstream sx ops.
- `inst.Store` carries `val_ty: TypeId` so the interp's raw-pointer
  store honours the destination width — no more "assume 8 bytes"
  silent clobber. Regression test at
  `examples/132-comptime-typed-store-widths.sx` exercises every
  primitive width (u8/u16/u32/u64, s8..s64, bool, f32, f64) via
  comptime checksums compared to runtime checksums.
- Call-convention mismatch at bare-fn → fn-pointer coercion is now
  a compile error (commit `f886d5f`). The chess-debug sweep that
  surfaced the bug also moved `#foreign` decls to default
  `callconv(.c)` and fixed every consumer-side sx callback exposed
  to a C API. Regression test:
  `examples/131-callconv-mismatch-diagnostic.sx`.
- Interp silent-arm sweep (commit `e9df33a`). Every `else =>` arm
  has a named bail reason via `bailDetail` / `typeErrorDetail`.
  `.deref` and `.unbox_any` used to silently pass through arbitrary
  Value kinds — now enumerated. `#run const` errors no longer
  swallow into `void_val`; emit_llvm surfaces them via std.debug.
- C-side callback into sx requires `callconv(.c)` on the sx fn (and
  on any fn-pointer TYPE the user casts a C fn-pointer through).
  Tests adjusted: `examples/61-objc-roundtrip.sx`,
  `examples/62-objc-class.sx`, `examples/95..97-objc-block*.sx`,
  `examples/ffi-06-callback.sx`.

154/154 example tests pass (two new regression tests added: 131 and
132). Chess green on macOS / iOS sim / Android.

ISSUE-MEM-002 (the `context.allocator.alloc(size)` pattern-match
bypass) is FULLY CLOSED. User-typed `context.allocator.X` flows
through the real protocol vtable at codegen *and* runs the same
chain at comptime in the interp. No remaining shortcut.

## Next step

Phase 1.3 (closure env allocation through context) shipped in commit
`8e21cc5`. Phase 1.4 (codegen serializer for all interp Value
variants) shipped this session. Phase 1.2 (free / malloc through
context) was considered and **skipped** — `context.allocator.alloc
/dealloc` already works directly; wrapper-only `malloc`/`free` would
be lossy renames.

The `List(T)` allocator-arg work shipped earlier this session is
**outside the original MEM plan** but lives in the same problem
space (long-lived container growth silently capturing the wrong
allocator).

Open follow-ups, in roughly the order they make sense:

- **`.int → ptr` heap-walk follow-up.** Phase 1.4a handles the
  fat-pointer aggregate case. A `.int` host-address landing in a
  bare ptr field (e.g. a struct with `*Inner` or `[*]u8` members)
  still bails. Investigated this session — the practical blocker
  is interp-level: `struct_gep` / load / store through raw integer
  pointers aren't wired (only `index_gep` is), so users can't even
  construct most cases. Two-step plan if a real trigger surfaces:
  (1) lift the interp's struct_gep+load+store-through-raw-int limit
  using `FieldAccess.base_type` for the field offset table; (2) add
  recursive heap-walk in `valueToLLVMConst` with cycle detection on
  `(addr, type_id)` visited pairs. No practical trigger in-tree —
  the canonical buffer/string case is already handled by `[]T` /
  `string`.

## Phase 0.3 audit findings — chess allocator usage (closed)

After Step 5 / matchContextAllocCall removal, every consumer call
to `context.allocator.X` flows through the real protocol vtable.
This section is left for history — the audit drove which sites
needed migration, but no chess code actually needed any allocator-
API change. The sites that used to bypass the protocol via the
`.heap_alloc` pattern-match now dispatch through the inline
Allocator value naturally.

- `~/projects/game/main.sx` — 7 sites of
  `context.allocator.alloc(size_of(T))` for platform/GPU/pipeline
  state. Now real protocol dispatch.
- `~/projects/game/chess/game.sx` — `ChessGameState.init` captures
  `context.allocator` into a `parent_allocator: Allocator` field
  and restores it via `push Context.{ allocator = ... }`. Worked
  before, still works.
- `~/projects/game/chess/pieces.sx` — declares its own `free`
  bound to libc and calls it on a libc-malloc'd buffer (from a
  foreign reader). Intentional C-interop bypass — no change needed.
- `~/projects/game/quick.sx` — quicksort demo. Same flow as main.sx.

## Log

- **2026-05-25 (latest)** — `resolveType(null) → .s64` removed.
  Signature changed to non-optional `*const Node`; 12 callers
  surveyed and classified. The three unguarded ones — top-level
  `var_decl` at `lower.zig:630` (now mirrors lowerVarDecl's
  infer-from-initializer pattern), `#objc_call` / `#jni_call` return
  types (already non-optional in AST so no actual silent fallback
  was reached) — handle null explicitly. The rest were guarded by
  `if (x != null)` blocks; cleaned up to optional-payload syntax.
  `examples/137-toplevel-var-type-inference.sx` proves the visible
  win: `g_pi := 3.14;` at module scope now infers `f64` (used to be
  silent `s64`). 159/159 + chess clean.
- **2026-05-25 (penultimate)** — Phase 1.4a shipped. `valueToLLVMConst`
  takes IR `TypeId` (not LLVM type) + an interpreter handle.
  String/slice fat pointers are serialized by capturing the
  pointed-to bytes (via `interp.heapSlice` for heap_ptr, raw
  process memory via new `readHostBytes` for byte_ptr / .int /
  string literal) and emitting a private global byte array. Struct
  / array aggregates recurse with declared field/element TypeIds.
  The trap case (`.int` landing in a ptr slot outside a fat
  pointer) bails loudly. `Interpreter.heapSlice` promoted to
  `pub`. Regression: `examples/136-comptime-string-global.sx`.
  158/158 + chess green on all three platforms.
- **2026-05-25 (penultimate)** — Allocator `init` returns the state by
  value. GPA / Arena / TrackingAllocator all changed; `Arena.deinit`
  no longer self-deallocs. `UIPipeline.arena_a/_b` embedded as values;
  `@self.arena_a` at the *Arena use site. `examples/50-smoke.sx`
  Arena alloc counts dropped by 1 (no state struct alloc); FFI IR
  snapshots regenerated to reflect new signatures (`-> ptr` →
  `-> i64`/`-> void sret(...)`). CLAUDE.md "Allocator construction"
  rewritten around the by-value convention. 157/157 + chess green
  on all three platforms via `tools/verify-step.sh`.
- **2026-05-25 (penultimate)** — `xx <lvalue>` semantics changed to borrow.
  Single change at `lower.zig:10334` (`buildProtocolErasure`) gated by
  new `isLvalueExpr` helper at `lower.zig:10322`. specs.md §3
  ownership table extended (three modes: rvalue / lvalue / pointer).
  `examples/130-xx-value-routes-through-context-allocator.sx` updated
  to use a struct literal (rvalue) as the operand — the heap-copy
  routing through `context.allocator` is what Phase 1.1 actually
  proves, and that path is still active for rvalues. New regression
  at `examples/135-xx-lvalue-borrows.sx` witnesses the borrow path
  via TrackingAllocator counts on the local. 157/157 + chess green
  on all three platforms (`tools/verify-step.sh` ran green
  immediately before this).
- **2026-05-25 (penultimate)** — Phase 1.4 shipped. `valueToLLVMConst`
  (`emit_llvm.zig:734`) replaced the primitive-only switch with a
  full serializer covering null_val, void_val, undef, func_ref,
  string, and aggregate (struct + array via
  `LLVMStructGetTypeAtIndex` / `LLVMGetElementType`). Unsupported
  variants (heap_ptr, byte_ptr, slot_ptr, closure, type_tag) bail
  loudly via `std.debug.print` with the global name. The call site
  at line 676 now passes `global.name` so the diagnostic locates the
  offending `#run` site. `lowerComptimeGlobal` (`lower.zig:6384`)
  infers the return type from the expression when the user omits
  the type annotation — closes the silent-s64 default for `NAME ::
  #run expr;` bindings. The broader `resolveType(null) -> .s64`
  fallback is left in place for other callers — flagged for a
  follow-up audit. Regression at
  `examples/134-comptime-aggregate-global.sx`. 156/156 + chess green.
- **2026-05-25 (penultimate)** — `List(T)` mutation API gained an optional
  trailing `alloc: Allocator = context.allocator` argument
  (`library/modules/std.sx`). Default-arg substitution previously
  only fired for identifier callees; extended to the generic-method
  dispatch path via new `appendDefaultArgs` helper at
  `lower.zig:7974-7991`, wired in at `lower.zig:5332`. Long-lived
  owners that grew internal Lists during render — `GlyphCache`,
  `DockInteraction`, `StateStore`, `Gles3Gpu`, `MetalGPU` — now
  capture `parent_allocator: Allocator` at init and forward it to
  every internal `.append` / `.alloc` / `.dealloc`. Chess panel-text
  regression (text vanished after the first move because GlyphCache
  hash + entries grew into the per-frame arena and died on reset)
  fixed end-to-end on macOS. specs.md §11 gains a "Default Parameter
  Values" subsection documenting the existing capability. Operator-
  precedence fix kept in `pipeline.sx` (`(self.frame_index & 1) == 0`
  instead of `self.frame_index & 1 == 0`, which was parsing as
  `self.frame_index & (1 == 0)` = always 0). All diagnostic logging
  added during the bug hunt has been stripped. 155/155 example tests
  green.
- **2026-05-25 (late)** — Interp silent-arm sweep (`e9df33a`).
  Every `else =>` arm has a `bailDetail` reason; `.deref` /
  `.unbox_any` previously silently passed through arbitrary Value
  kinds, now enumerated. `#run const` errors surface a real
  diagnostic via emit_llvm instead of becoming `void_val`.
  CLAUDE.md REJECTED PATTERNS gained the "silent unimplemented
  arms" section (`4de565b`). 154/154 + chess green.
- **2026-05-25 (mid)** — Typed raw-pointer stores (`f2b3868`).
  `inst.Store` carries `val_ty: TypeId`, threaded by
  `builder.store` and consumed by `storeAtRawPtr` to write exactly
  the declared destination width. Regression at
  `examples/132-comptime-typed-store-widths.sx` exercises every
  primitive width via comptime/runtime checksum comparison.
  `index_get` raw-pointer arm added (was bailing). Comptime init
  errors no longer swallow into zero.
- **2026-05-25 (mid)** — Drop `matchContextAllocCall` (`d415bcc`).
  Comptime now runs the full Allocator-protocol dispatch chain by
  reusing the parent module instead of spinning up a fresh
  ct_module. Interp gained `.int` / `.byte_ptr` arms in
  `index_gep`, `store`, `marshalForeignArg`, `asString`. Closes
  ISSUE-MEM-002 fully. JNI stub binding extended to call
  current_ctx_ref → &__sx_default_context (used to be gated on
  isExportedEntryName).
- **2026-05-25 (mid)** — Reject call-conv mismatch at bare-fn →
  fn-pointer coercion (`f886d5f`). `#foreign` decls now default to
  `callconv(.c)`. Library audit (`619aff8`) — all C-side callbacks
  already followed the rule; documented the one remaining gap
  (`xx <sx_fn> : *void` cast to opaque, ambiguous from cast alone).
  Regression at `examples/131-callconv-mismatch-diagnostic.sx`.
- **2026-05-25 (early)** — Implicit-Context refactor SHIPPED
  end-to-end. All 9 plan steps
  (`lets-see-options-for-merry-dijkstra.md`) landed. `context` is
  no longer an LLVM global; every sx function carries `__sx_ctx`
  at slot 0; `context.X` reads load through `current_ctx_ref`;
  `push Context.{...}` is alloca + rebind; FFI-inbound entries
  install `&__sx_default_context`; interp bootstraps the default
  Context on top-level call. 152/152 + unit tests green.
  Commits: `29784c2` (Steps 1-2), `92c6b47` (Step 3),
  `4bf5908` (Steps 5-7), `b69a2ea` (Step 8).
- **2026-05-24** — Phase 1.1 shipped: `buildProtocolValue` heap-copy
  now routes through `context.allocator.alloc` via the new
  `allocViaContext` helper. Regression at
  `examples/130-xx-value-routes-through-context-allocator.sx`
  proves a `Tracer` installed via `push Context` sees the alloc
  (`Tracer.count = 1`) — interp + codegen parity. 152/152 +
  chess green.
- **2026-05-24** — issue-0041 and issue-0042 both fixed end-to-end.
  Also removed the silent `.s64` fallback in `resolveTypeArg`,
  guarded the two upstream callers (`buildTypeBindings`,
  `inferGenericReturnType`) with `type_bridge.isTypeShapedAstNode`,
  and fixed three parser regressions introduced by the 0041 work
  (`[:0]u8` return type, nested return-arrows, `struct #compiler`
  trailing-`;` method decls). Full verify-step.sh gate green.
  CLAUDE.md REJECTED PATTERNS gained the silent-fallback-defaults
  prohibition. Stream now READY for Phase 1.
- **2026-05-24** — Phase 0.6 shipped (`align_of($T)` builtin).
  Touchpoints: `inst.zig` BuiltinId, `lower.zig` registry +
  return-type table + reflection handler, `interp.zig` fallback,
  `sema.zig` builtin allowlist, `lsp/server.zig` both completion
  tables, `library/modules/std.sx`. Smoke coverage added in
  `examples/50-smoke.sx`. 151/151 + chess green on all platforms.
  Then `size_of(*u8)` parse error was investigated — filed as
  issue-0041 (pre-existing, affects both `size_of` and `align_of`).
  Stream paused on 0041. Also tightened CLAUDE.md IMPASSIBLE RULES
  to close the "pre-existing / non-blocking" loophole that almost
  let this session roll past the issue filing.
- **2026-05-24** — issue-0040 filed. Phase 0.9 verified that
  `obj.method(Type)` with a `$T: Type` parameter fails to dispatch
  via dot, while explicit static (`T.method(obj, Type)`) and pipe
  (`obj |> T.method(Type)`) both work. Root cause pinpoint:
  `src/ir/lower.zig:5066-5123` has branches for generic-template
  struct methods (5082) and non-generic qualified (5106), but no
  branch for a non-template struct with a generic method. Stream
  paused on 0040.
- **2026-05-24** — Phase 0.8 `#foreign("c")` syntax verified.
  Form is `name :: <sig> #foreign libc ["c_name"]` with the optional
  string literal supplying a rename. Confirmed via
  `libc_strlen :: (s: *u8) -> usize #foreign libc "strlen";`
  scratch test (interp + codegen parity).
- **2026-05-24** — issue-0039 fix verified, full
  `tools/verify-step.sh` gate green again: 150/150 example tests
  pass and chess builds + screenshots OK on macOS / iOS sim /
  Android.
- **2026-05-24** — issue-0038 fix verified; spike now errors as
  expected on transitive references. 149/149 example tests pass
  (+1 vs pre-fix). Chess build broken as predicted — three-bucket
  triage written up in `issues/0039-chess-needs-explicit-imports-
  post-0038.md`. Stream paused on 0039.
- **2026-05-24** — Phase 0.7 spike at `/tmp/sx-import-spike/`
  (`a.sx → b.sx → c.sx`) showed `a.sx` calls `c_only_fn()` and reads
  `c_only_const` directly. Filed as issue-0038. Stream is paused per
  the IMPASSIBLE RULE — no workaround, the libc hide-by-internal-
  module strategy in the plan depends on the language semantics
  matching the assumption.
- **2026-05-24** — Phase 0.2 sanity sweep landed. Five additional xx
  cast patterns exercised through `tools/scratch.sh`; all show
  interp/codegen parity. Combined with the issue-0037 regression at
  `examples/126-xx-recover-then-dispatch.sx`, the cast story is now
  considered tight for the cases the MEM plan relies on.
- **2026-05-24** — Phase 0.3 chess allocator audit recorded above.
  No chess-side migration needed; ISSUE-MEM-002 (`context.allocator`
  bypass) is the only thing the chess codebase is exposed to and
  Phase 1 already owns it.
- **2026-05-24** — Phase 0.0a (`tools/verify-step.sh`) shipped.
  Confirmed working: 145/145 example tests + chess builds + screenshots
  on all 3 platforms. Initial 3-second screencap delay was too short
  for Android — increased to 6 seconds; iOS sim + macOS to 5 seconds.
- **2026-05-24** — Phase 0.0b (`tools/scratch.sh`) shipped. Verified
  with a hello-world snippet: interp + codegen agree.
- **2026-05-24** — Phase 0.0c initial implementation of
  `TrackingAllocator` in `library/modules/allocators.sx`. Build + 145
  tests pass after snapshot regen. Chess green on all 3 platforms.
  Manual scratch.sh test confirms counters increment correctly when
  called directly on the tracker variable.
- **2026-05-24** — ISSUE-MEM-007 fixed in `src/ir/lower.zig`.
  Root cause: `emitProtocolDispatch` keyed the auto-unbox path on
  `mi.ret_type == void_ptr`, but `*void` is overloaded — both
  Self-disguised-as-*void AND a literal `-> *void` return appear as
  the same `TypeId`. With `target_type` leaking from the enclosing
  function's return type (e.g. `main -> s32`), every `*void` return
  was loaded as `s32`, yielding 0 → null. Fix: stash `ret_is_self`
  on `ProtocolMethodInfo` during `registerProtocolDecl` (set when
  the AST return-type node is the `Self` type-expr), and gate the
  unbox on that flag. Regression at
  `examples/99-protocol-void-pointer-return.sx`. Sister symptom
  (SIGTRAP inside struct-static method storing an Allocator field)
  also fixed by the same change.
- **2026-05-24** — issue-0037 fixed in `src/ir/lower.zig`. Root
  cause: `lowerXX` had a Concrete→Protocol erasure branch but no
  inverse Protocol→pointer recovery — the cast fell through to
  `coerceToType`, which couldn't match the (struct, pointer) shape
  and returned the operand unchanged. Result: a 24-byte protocol
  struct stored into an 8-byte ptr alloca, corrupting adjacent
  stack (the protocol value's own slot was the next victim, so the
  next dispatch loaded garbage and crashed). Fix: when src is a
  protocol value and dst is a pointer, emit `struct_get` of field 0
  (ctx), then bitcast `*void` → dst. Regression at
  `examples/126-xx-recover-then-dispatch.sx`.

## Known issues (discovered during execution)

### ISSUE-MEM-001: Type inference defaults `p := malloc(64)` to `s64`

**Severity:** medium (workaround exists; bites unexpectedly).

**Symptom:** Writing `p := malloc(64)` (no explicit type) infers
`p: s64` instead of `p: *void`. Subsequent `free(p)` then fails LLVM
verification with "Call parameter type does not match function
signature!" because `free` expects `ptr` but receives `i64`.

**Workaround:** Explicit type annotation `p : *void = malloc(64);`
or `xx malloc(64);` at the call site.

**Reproduction:**
```sx
main :: () -> s32 {
    p := malloc(64);   // p inferred as s64
    free(p);            // LLVM verify fails: ptr expected, i64 given
    0;
}
```

**Root cause:** Likely in the inference path for `:=` declarations
when the RHS is a `*void`-returning #builtin. The compiler defaults
the binding to s64 instead of matching the return type. To
investigate in a future session.

**Status:** Open. Not blocking mem.sx work but worth fixing as a
quality-of-life issue. File as `examples/issue-NNNN.sx` when
addressed.

### ISSUE-MEM-002: `context.allocator.alloc/dealloc` bypasses protocol dispatch

**Severity:** high (breaks the entire Phase 1 premise; documented
fix in Phase 1).

**Symptom:** Any code that goes through `context.allocator.alloc(size)`
or `context.allocator.dealloc(ptr)` is pattern-matched in
[src/ir/lower.zig:5137-5159](src/ir/lower.zig#L5137-L5159) and lowered
directly to `.heap_alloc`/`.heap_free` IR — which calls libc
malloc/free. The protocol-value vtable is bypassed entirely.

This means a `push Context { allocator = my_tracker }` block
followed by `context.allocator.alloc(size)` does NOT call the
tracker's `alloc` method. The tracker sees zero allocations even
though many occurred.

**Workaround:** Call the allocator directly via a variable rather
than via `context.allocator`:
```sx
push Context.{ allocator = tracker, data = null } {
    p := tracker.alloc(64);     // works — dispatches through tracker
    tracker.dealloc(p);
}
```

But this defeats the purpose of context-allocator overriding for
user code that doesn't know about the tracker.

**Fix:** Phase 1 of the mem.sx plan removes this pattern-match and
replaces it with proper context dispatch through the Allocator
protocol. After Phase 1, `context.allocator.alloc(size)` correctly
dispatches to whatever allocator is currently set in context.

**Status:** Documented in plan; fixed in Phase 1. Blocks
auto-tracker-wrap (Phase 5 `--leak-check`).

### ISSUE-MEM-003: 08-types test depends on undefined memory contents

**Severity:** low (flaky test exposes itself when allocator code
changes).

**Symptom:** `examples/08-types.sx` declares a struct field
`c : u8 = ---;` (uninitialized) and prints the struct. The expected
snapshot captures a specific value (formerly `c: 176`, now `c: 8`)
which depends on whatever's in undefined memory at that moment. Any
allocator change shifts the value.

**Workaround:** Regenerate snapshot via `bash tests/run_examples.sh
--update` when this test fails for unrelated reasons.

**Fix:** Test should be rewritten to NOT depend on undefined memory
content — perhaps verify the field is one of N specific values, or
just don't print uninitialized data. Out of scope for mem.sx.

**Status:** Open. Document and live with it for now.

### ISSUE-MEM-004: Adding code to allocators.sx shifts JNI/Obj-C IR snapshots

**Severity:** medium (eats time per step).

**Symptom:** Every additive change to allocators.sx (new struct,
new method) cascades into ~11 IR snapshot diffs across the
`tests/expected/ffi-{jni,objc}-*.ir` files. Each step needs
`--update` + `git diff` review. The diffs are usually benign
(additive declarations) but one (`ffi-objc-call-06-sret-return.ir`)
reorganises ~1500 lines because it uses reflection on the full type
registry.

**Workaround:** Per-step `--update` + diff review.

**Fix:** Phase 0.1 spike — extend `normalize_ir()` in
`tests/run_examples.sh` to strip allocator-related declarations and
constant-pool renumberings from snapshots. Should make additive
changes invisible at the snapshot level while preserving JNI/Obj-C
ceremony signal.

**Status:** Mitigated; structural fix in Phase 0.1.

### ISSUE-MEM-005: Two-line `create(@storage)` pattern for allocators

**Severity:** low (cosmetic but real DX friction).

**Symptom:** Current pattern requires a separate storage decl + a
create call:
```sx
g_gpa : GPA = ---;
libc := GPA.create(@g_gpa);
```

Two lines per allocator. Plan committed to one-line via heap-copy:
```sx
libc := GPA.create();   // heap-copies via xx value
```

**Fix in progress:** Phase 0.0c — switch `create` methods on at
least TrackingAllocator (and possibly GPA/Arena/BufAlloc) to use
`xx value` heap-copy. Add `instance(a) -> *T` accessor for types
where users need the underlying state (TrackingAllocator).

**Status:** In progress (active work).

### ISSUE-MEM-007: Protocol dispatch on `*void`-returning methods returns null (FIXED)

**Severity:** was CRITICAL; resolved.

**Root cause:** `emitProtocolDispatch` in `src/ir/lower.zig` gated
its auto-unbox path on `mi.ret_type == void_ptr`, but the same
`TypeId` covers both Self-disguised-as-*void and a literal
`-> *void`. With `target_type` leaking from the surrounding
function (e.g. `main -> s32`), every protocol call returning
`*void` got its result loaded as `sizeof(target_type)` bytes — for
`s32` that's the first 4 bytes of the malloc'd block, which were
zero, comparing equal to null.

**Fix:** Tag `ProtocolMethodInfo` with `ret_is_self: bool`, set in
`registerProtocolDecl` when the AST return-type is the `Self`
type-expr. `emitProtocolDispatch` only takes the auto-unbox path
when `ret_is_self` is true. Literal `*void` returns are now passed
through unchanged.

**Regression:** `examples/99-protocol-void-pointer-return.sx`. The
sister symptom (SIGTRAP from protocol dispatch inside a struct
static method that stores an Allocator field) was the same root
cause and is also fixed.

**Status:** Resolved 2026-05-24.

### ISSUE-MEM-006: Android screencap needs 6+ seconds; iOS sim + macOS 5+

**Severity:** low (fixed in tooling).

**Symptom:** Initial `verify-step.sh` used `sleep 3` before
screencap. Android needed longer for the side panel to render;
without it, the screenshot showed only the chess board and a black
strip where the side panel should be.

**Fix:** Updated `verify-step.sh` delays: macOS 5s, iOS sim 5s,
Android 6s. Documented inline in the script.

**Status:** Resolved.