fix(smem-hygiene): default the Stop hook's session-summary fallback to off

[RobertSigmundsson]

Summary

• The Stop hook's session-summary fallback becomes opt-in: auto.capture_session_summary, defaulting to false.
• The fallback is a truncation, not a summary; on a stock installation the write gate is off, so what it produces is stored rather than filtered.
• Adds six tests over the default, the round trip, and the behaviour on both settings.

Why

When pattern extraction finds nothing in a transcript, the Stop hook manufactures a candidate from the last few lines and offers that to the write gate. On a real chat transcript those lines are harness markers and half-sentences, not knowledge — so the gate rejects them, and the work of extracting, encoding and judging them is repeated on every turn that produces no patterns.

The repository already says as much in two places, which is what makes this a default worth changing rather than a preference:

• quality_scorer.py's _NOISE_PREFIXES opens with the literal "session activity:", above a comment describing it as an "Empirical denylist from gate_decision false-accepts (2026-07-23): 180x "Session activity:", 30x
  , …". The single most common false accept the gate had to learn to reject is the thing this fallback manufactures.
• capture_state.py's module docstring calls duplicate "Session activity" entries "the dominant source of memory poisoning".

Measured on one production brain over twenty-four hours: 96 of 158 write-gate rejections, a little over three in five, originated in this path alone — entries like Session activity: <task-notification>. That number is from a single brain and I would not lean on it alone; it agrees with the direction of the two notes above, which are the project's own.

Changes

• unified_config.py: the new auto.capture_session_summary field, defaulting to false, with the reason recorded beside it.
• hooks/stop.py: the fallback runs only when the flag is set.
• tests/unit/test_stop_hook_session_summary.py (new) and tests/unit/test_hook_capture_idempotency.py, which turns the fallback back on explicitly.
• docs/reference/config.md: regenerated with scripts/gen_config_docs.py, since that file is generated from these defaults and the Docs Freshness job checks it.

What this does not change

The gate itself. WriteGateConfig defaults to enabled: False and mode: "off", so on a stock installation the fallback's output is not rejected — it is stored, as a CONTEXT memory. That is the case this default actually addresses, and it is the common one rather than the exception, which is worth being explicit about: the argument here is not "the gate wastes effort" but "on the defaults, nothing is stopping this".

Setting capture_session_summary = true restores the previous behaviour exactly. The #80 idempotency regression tests do that deliberately, since the defect they pin lives on this path and needs it switched on to be observable at all.

Not touched, on purpose

The extraction helper keeps its name. Renaming it to say "tail" rather than "summary" would describe it better, and it has a single call site, so the rename would be small — but a rename is a different change from a default flip and belongs in its own commit rather than riding along.

Test plan

• pytest tests/unit/test_stop_hook_session_summary.py — 6 passed: the default is off, an absent key loads as off, an explicit true survives the config round trip, save emits the key, and the hook stores nothing for pattern-free text with the flag off and does store with it on.
• pytest tests/unit/test_hook_capture_idempotency.py — 12 passed. Those tests turn the fallback back on explicitly, because the Auto-capture: no idempotency key (Stop/PreCompact re-capture loop) + PreCompact/task_context hooks still bypass dedup #80 defect they exist to pin is only reachable through it.
• With hooks/stop.py and unified_config.py reverted to main and the tests kept, 5 of the 6 new tests fail.
• pytest tests/ -m "not stress" -n 4 — 7289 passed, 48 skipped, 1 xfailed, which is main's 7283 plus the six 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 740 files already formatted.
• mypy src/ --ignore-missing-imports — success, no issues found in 354 source files.
• Coverage under the CI gate: 72.37%, 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.

Verified by

@RobertSigmundsson

[acidkill]

Reviewed the full diff. The measured 61%-of-rejections number justifies the default flip; both directions are pinned, the idempotency suite opts back in explicitly, and the subprocess racer gets config.toml rather than an in-process toggle. Note for the changelog: default behaviour change, opt-out is auto.capture_session_summary=true.