Detect a torn packed-cache read instead of serving wrong rows

[thorwhalen]

Detects a torn packed-cache read instead of silently serving mismatched rows.

CorpusStore._save_packed writes matrix.npy, ids.json, metas.json and
sig.json as four independent calls. Two concurrent rebuilds of a same-size
corpus can leave the matrix from writer A beside ids/metas from writer B; the
old _load_packed only checked len(ids) == mat.shape[0], so the mixture
passed every check while row i no longer belonged to ids[i] — searches
answered confidently wrong, nothing raised or logged.

sig.json now also carries a sha256 content_sig over the exact ids/metas
bytes written with that matrix, plus the matrix shape. On load, a
disagreement is treated as a cache miss and the matrix rebuilds from records
— the same fallback path already used for a missing sig, bad format, or
OSError.

This is option 2 of #77 (detect cheaply and loudly), not option 1 (atomic
directory-pointer swap); a store-level lock / true write atomicity remains
open there, so this closes #77 for the corruption-detection half of it.

Backward/forward compatible: checks are guarded on the new fields being
present, so a cache written before this loads on the length checks alone, and
_PACKED_FORMAT stays at 1 (no cache invalidation). No public name,
signature or return type changes — every packed helper is module-private,
and stores opened without packed_dir (e.g. CorpusStore.memory()) never
reach this code.

What I did (rebase + gate)

• Branch pushed 2026-09-07, no PR opened. Rebased cleanly onto current
  master (2 commits it was missing: Replace argh with cw, preserving the BY_NAME_IF_KWONLY grammar #79, feat: a records corpus source whose records another package owns #80) — no conflicts.
• wads ci-local: ruff format + lint, pytest on py3.10 and py3.12 (556
  passed, 5 skipped), build — all green.

Dependents check

fleet_dependents.json lists raglab and truffle as importers of ir.

• raglab (present on this box) only calls CorpusStore.memory() and the
  public CorpusStore surface in its tests/source — it never sets
  packed_dir and never touches _load_packed/_save_packed, so it cannot
  reach this code path at all (confirmed by grep, matching the commit's own
  claim that stores without packed_dir never reach this code).
• truffle is not cloned on this box, so it could not be checked directly.
  Given the change touches only module-private packed-cache helpers with no
  public signature/return-type/name change, this is not treated as a
  breaking change requiring truffle's tests — noted here for the record.

Closes #77

🤖 Generated with Claude Code

[thorwhalen]

Post-merge refute review: content_sig covered ids/metas but not the matrix. The interleaving "A saves matrix, B saves a full set, A writes ids/metas/sig" still served B's rows under A's ids. Fixed in #83 (per-writer generation files + atomic sig publish), with a regression test that drives the real interleaving.