# Session summary — corpus sweep (2026-10-18)

Branch **`resolve-check`**, tree clean.  Suite **223/223**; fixpoint
**OK (3,731,524 bytes)**; corpus **382/612 compile-OK, 0 crashes**.

Tags: `v3-builtins-odd-trunc-float`, `v3-bool-enum-array-bounds`,
`v3-set-of-enum`, `v3-fwd-set-of-type`, `v3-corpus-382`,
`v3-corpus-sweep-2026-10-18`.

## What this session did

Continued the corpus-driven language work (the top backlog item from
`docs/STATUS.md`): re-measured, then fixed the bounded gaps one at a
time, each with a suite test.  Corpus **372 → 382** (+10 files), 0
crashes.

| Commit | Tag | Change | Corpus |
| --- | --- | --- | --- |
| `368b20d` | `v3-builtins-odd-trunc-float` | `ODD(x)` → `(x&1)<>0`; `TRUNC`/`FLOAT` as `VAL` sugar | 372 → 376 |
| `0c7fa60` | `v3-bool-enum-array-bounds` | `BOOLEAN`/`ENUM` array bounds + indices (`BoundPair` accepts `ClBool`/`ClEnum`; 218 accepts a BOOLEAN index) | 376 → 381 |
| `d76822a` | `v3-set-of-enum` | `SET OF <enum>` literals (`BaseSpanOk` accepts an enum base) | 381 |
| `1de6c16` | `v3-fwd-set-of-type` | forward type ref in `SET OF <type>` (`SymTab.Unresolved` + deferred `SetType` span check) | 381 → 382 |

New tests: `t_truncfloat`, `t_enumidx`, `t_setenum`, `t_fwdset`.

Side note: `t_bad_index` had asserted a *non*-error (BOOLEAN is a legal
ordinal index) — changed to use a `REAL` index.

## Findings recorded (not fixed — features, not one-liners)

`docs/analysis_gaps.md` §9:

1. **Set constants** (`CONST Digits = CharSet{"0".."9"}`).  Accepting the
   declaration is easy (treat `ClSet` as an aggregate in `ConstDecl` and
   the designator's `KindConst` branch — written, then **reverted**):
   the *value* needs **static** data, but a set literal is a **runtime
   temp** and `QbeGen.DeclSet` only emits a zeroed blob.  The fix is a
   parse-time mask emitted as `data $X = { w <mask>… }` + a `Lower`
   re-emit path.  Reverted to a clear `230` rather than miscompile.
2. **Builtin calls in `CONST`** (`CAP(VAL(CHAR,65))`, `MIN`/`MAX` of
   sets): `CAP`/`ABS`/`ODD` set the `"@"` sentinel so `ConstDecl`'s
   `IsImm` rejects them; fix is to fold builtins over immediates.

## Tooling / workflow

Switched scratch space to a **repo-local `.scratch/`** (gitignored,
never committed) because `/tmp/opencode` is wiped between days:
- corpus wrapper `./.scratch/corpusx.sh` (work dir `.scratch/v3corpus/`),
- repro modules (`.scratch/*.mod`),
- fixpoint staging via `TMP=$PWD/.scratch/fixpoint`.

Docs refreshed: `README.md`, `docs/STATUS.md` (numbers), `docs/INDEX.md`
(navigation), `docs/analysis_gaps.md` §8–§9.

## Resume

```sh
cd .../m2compiler-V3
git checkout resolve-check                  # HEAD d3aad6f
cd compiler && ./build.sh && ./run_tests.sh # 223/223
cd .. && TMP="$PWD/.scratch/fixpoint" ./bootstrap/fixpoint.sh

# corpus (local scratch):
M2FLAGS="" ./.scratch/corpusx.sh            # 382 OK, 0 crash
```

## Next

- **Set constants** (static mask) and **builtin CONST folding** — the two
  recorded features.
- Wider corpus buckets: `SET OF` over wide bases, `ValLabel`, `ImpName`,
  parser `ident/':'/END expected` cases; and the long-standing
  **size-sensitive V3 miscompile**.