fix(smem-hygiene): remember the gate's refusals, not only its saves

[RobertSigmundsson]

Summary

• The auto-capture hooks now record the write gate's refusals in the idempotency ledger, not only its acceptances, so a rejected fragment is judged once per session instead of on every Stop.
• PreCompact honours a refusal Stop already recorded, and records its own.
• Adds ten tests, including three that count gate invocations rather than trusting a save count.

Why

The ledger from #115 remembered what the gate accepted. A refusal left no trace at all, so the next Stop in the same session re-extracted the same fragment, re-encoded it, and put it to the gate again — which, being deterministic, refused it again. PreCompact then repeated the exercise from the top. Nothing crashes and nothing corrupts; the gate simply does the same homework on the same content forever, which is a shame in a mechanism that exists to stop precisely that.

Measured on one live brain over twenty-four hours on 2026-08-08: 499 auto decisions carried just 134 distinct contents, one fragment having been judged thirty-six times. That inflates the gate's denominator about 3.7-fold and makes its observed accept rate look several times worse than it is — 0.39% counted over rows against roughly 1.4% over distinct content. The wasted encoding is the smaller half of the cost; the misleading acceptance statistic is the half that gets acted upon.

Changes

• hooks/capture_state.py: vocabulary for recording a refusal, keyed so that it is scoped to the threshold that produced it. A refusal recorded under one threshold does not suppress a re-judgement after the threshold changes, whilst an acceptance survives such a change.
• hooks/stop.py: the main capture loop records the refusal alongside the outcomes it already recorded.
• hooks/pre_compact.py: honours a refusal recorded by Stop rather than re-deciding it, and records its own refusals the same way.
• tests/unit/test_capture_state.py (+6) and tests/unit/test_hook_capture_idempotency.py (+4).

One rough edge, disclosed rather than hidden

When every detected fragment was duplicate-filtered, the run reports the existing duplicate content already captured status and does not fall through to a session summary. Reusing that message is slightly imprecise — the two situations are not identical — but replacing an accurate duplicate status with a brand-new summary save would be the worse outcome, and a separate counter looked like more machinery than the distinction earns. Happy to split it if you would rather the two were distinguishable.

The module docstring in test_hook_capture_idempotency.py also loses a stale absolute path from a developer's machine. The file is being edited here anyway, so it seemed a good moment.

Test plan

• pytest tests/unit/test_hook_capture_idempotency.py tests/unit/test_capture_state.py — 43 passed.
• Differential control, and the reason it is reported in two layers: an earlier shape of test_pre_compact_honours_refusal_recorded_by_stop passed with the fix entirely removed, because PreCompact never reached the gate on that input and the assertion could not distinguish that from success. The tests now count gate invocations and assert the gate was called before asserting it was not called again.
  • Reverting only hooks/pre_compact.py: 2 failed, 14 passed — test_pre_compact_honours_refusal_recorded_by_stop and test_pre_compact_records_its_own_refusal.
  • Reverting hooks/pre_compact.py and hooks/stop.py: 3 failed, 40 passed — the two above plus test_stop_hook_does_not_rejudge_rejected_content.
  • Each test is therefore pinned to the file whose hunk it covers.
• Line coverage of the change itself, rather than of the modules: every executable line this PR adds to capture_state.py, pre_compact.py and stop.py is executed by the tests above — 12 statements, 0 uncovered. (The modules' own totals are low because they are large hook entry points with a great deal of unrelated code.)
• pytest tests/ -m "not stress" -n 4 against a live SurrealDB v3.2.0 — 7293 passed, 48 skipped, 1 xfailed, which is main's 7283 plus exactly the ten tests added here. Two tests in tests/unit/test_dashboard_brains_scope.py fail on this branch and on main alike: they want a live database and collide with one another under -n. Both pass when that file is run on its own.
• ruff check src/ tests/ clean; ruff format --check src/ tests/ reports 739 files already formatted.
• mypy src/ --ignore-missing-imports — success, no issues found in 354 source files.
• Coverage under the CI gate: 72.44%, against 72.36% on main.
• CHANGELOG.md untouched — left to the release entry, as with fix(storage): bind datetimes in time comparisons so they select by value #191–fix(memory): refresh content-derived fields on compress, restore, and refine #193.

Related issues

None filed. #115 introduced the ledger this extends; the refusal half was simply never part of it.

Verified by

@RobertSigmundsson

[acidkill]

Reviewed the full diff. Rejection keys namespaced by threshold are the right shape — remembered refusals without silencing content if auto_capture_min_score is later lowered; shared seen-set across Stop and PreCompact kept consistent.