# Session 2026-10-19 — summary

Branch **`master`** (tree clean).  Eleven commits, tagged individually
(`v3-generics-two-instances` … `v3-pim-io-libs`).  Plan source:
`handoff-2026-10-18-next.md` (P0–P5) + `handoff-2026-10-19-next.md`.

## Numbers

| gate | session start | end |
| --- | --- | --- |
| suite (`compiler/run_tests.sh`) | 239/239 | **249/249** |
| fixpoint (`bootstrap/fixpoint.sh`) | OK (3,812,771 B) | **OK (3,882,337 B)** |
| corpus (`tools/v3-corpus/corpus.sh`) | 405/612 | **467/612, 0 crashes** |

Reproduce:

```sh
cd compiler && ./build.sh && ./run_tests.sh          # 249/249
cd .. && TMP="$PWD/.scratch/fixpoint" ./bootstrap/fixpoint.sh
tools/v3-corpus/corpus.sh                            # 467 OK, 0 crash
```

## What landed

### P0 — generics double-instantiation (real bug) — `dc3b3c8`
One generic instantiated twice with different-family type arguments
failed on the second.  Root cause was **Lowering**, not the formal
binding: after `EndUnit`, `Lower` re-emitted each unit in `globScope`,
so `MaterializeGlobals` copied the unit's types there; a generic instance
leaked its `NodeRec` and the second instance's `POINTER TO NodeRec`
resolved the first instance's record via `ForwardType`.
Fix: `SymTab.EnterLowerScope`/`LeaveLowerScope` (re-emit each unit in its
own scope), `Lower.ReenterImports` for nested-module FROM-imports.
Regression `g_twice*` / `GTwiceProg`.

### P1 — two gm2 extensions (user chose compatibility) — `a33fc7c`, `b042be3`
- `__FILE__` / `__LINE__`: scanner allows a leading `_`; both are
  reserved literals folded to the file-name string / line-number integer.
- scalar actual → `ARRAY OF BYTE`: parse-time marker on the open formal;
  a scalar is wrapped in a stack descriptor `[count=SIZE(actual)][copy]`
  (`QbeGen.OpenScalarDesc`).  Approximate by design (writes hit the copy;
  `BYTE` stays 32-bit) — documented in `gaps/README.md`.

### P2 — structural gm2 extensions — `81b488b`
`gap10` same-size scalar `VAR` actual; `gap11` PIM `PROC` type + bare
proc-value statement; `gap12` variant empty arms / `ELSE` / comma labels
/ untagged inner `CASE` / nested variant `CASE`; `gap18` set-typed `CASE`
labels (word-wise `CmpSet`).  Corpus +16.

### P3 — strict declaration order in bodies — `10e0951`
Deferred bodies are now order-aware for this module's own **data**
(variables/constants/types): `SymTab` declaration sequence + a window in
`Lookup`; parser state moved so the parser obeys the rule.  Forward
**procedure** calls stay allowed (the generated parser relies on it;
gm2's bootstrap link rejects `FORWARD`).  Corpus −6 (intended: gm2
`decl-order` leniencies now correctly rejected).

### P4 — the "size-sensitive" miscompile — `a50fc6f` (docs)
**Not size-sensitive, not a codegen bug.**  `ASTDUMP` dumped the AST to
stdout; built from its own sources that dump contained the driver's
literal `"Parsed correctly"`, and the fixpoint gate was `grep -q Parsed`
— so the gate passed while the session reported `Incorrect source`.
Removing the dump unmasked real errors (Lower's forward procedure calls),
already fixed in `a3ef897`.  Also corrected the stale note about the
record-array-field codegen bug (`38e0b54`): it is **fixed** (hybrid
`ArrBodyItems`), covered by `t_arrrecfld`/`t_arrrecbig`/`t_nestarr`.

### gap14 + P5 — `f53cdad`
- `__BUILTIN__` / `__ATTRIBUTE__` declaration pragmas parsed and ignored
  (`builtin.def/.mod` + 4 importers now compile).
- `docs/INDEX.md` rewritten; 45 unreferenced historical files moved to
  `docs/archive/` (all 69 cross-references kept valid).
- `gap13` remainder (standard built-ins as identifiers) documented and
  **deliberately deferred**: no corpus exemplar, and it needs a
  contextual-keyword rewrite of `Fact`/`ArgList` (Coco/R is LL(1)); a
  declaration-site-only measure would silently bind calls to the
  built-in.

### Robustness — `3d58325`
Table-limit overflows are now loud (`230`), not silent drops: the
two-phase span table (used to silently drop a body's codegen — regression
`t_manyproc`), the label/`GOTO` tables, the 32-actual AST buffer, and
`SymTab.MaxTypes`.  Corpus unchanged (no corpus file hits a limit).

### Library — `c7d5cff`, `d243d25`
- ISO 10514-1 `ChanConsts` + `StreamFile` on `IOChan`; PIM/gm2
  `NumberIO`, `RealInOut`, `FpuIO` (new shim `m2writefix`).
- Two real compiler bugs surfaced and fixed: **cross-unit set `CONST`s**
  (emitted under the global name now) and **qualified enum literals in
  set constants** (`Lower.EvalOrdConst`).
- Harness fidelity: `corpus.sh` now resolves V3's own shipped libraries
  by default (`V3_LIBS`), instead of mislabelling their imports `LIB`.

## Corrections / lessons

- The "size-sensitive miscompile" and the "still-open record-array-field
  bug" were both stale notes; both are resolved.  Verify old findings
  before acting on them.
- A weak success gate (`grep -q Parsed`) can mask real errors — the gate
  now greps `Parsed correctly` and fails on session errors.

## Remaining (for a next session)

1. **Corpus `LANG` bucket** — after the library work: `undeclared
   identifier` (gm2-only modules `M2RTS`/`Builtins`/`Processes`/… aside),
   `not supported yet`, `invalid call`.  Triage the top kinds.
2. **Error recovery / message polish** — dropped tokens cascade.
3. **`gap13` refactor** (optional, no corpus exemplar).
4. `STATUS.md` "what remains" is the live list.

Nothing was pushed; `origin/master` is the previous commit.