feat(write-gate): reject machine output and session-notification noise (Gate 1b)

[RobertSigmundsson]

Problem

Passive auto-capture stores a lot of harness output as if it were knowledge. Measured over one day of
gate_decision false-accepts on a live brain: 180× "Session activity:", 30× <summary>,
20× "Background command", 19× <status>, 12× task-notification, plus tool-use-id,
*-caveat> and raw shell lines (ps -o …, echo …, curl -…). None of it is a memory. All of it
passes the length and quality-score gates because it is long, structured, and full of distinct tokens.

Change

A new Gate 1b in score_memory, running before the scoring heuristics:

def _is_machine_noise(content: str) -> bool:
    """Two signals: (a) a known auto-capture/shell noise prefix, or
    (b) a structural notification/tool tag that never appears in a real memory."""

A match returns QualityResult(score=0, quality="low", rejected=True, rejection_reason="Machine-noise rejected: …") — same shape as the existing rejection gates.

Why upstream benefits

The denylist is derived from measured false-accepts, not guesswork, and it targets exactly the content
this project's own hooks generate. Every entry is a structural marker (a tag or a fixed prefix), so the
false-positive surface on genuine prose is small.

Evidence

Check	Result
cherry-pick onto origin/main	clean
import surreal_memory	OK
ruff check	All checks passed!
ruff format --check	701 files already formatted
mypy src/ --ignore-missing-imports	Success: no issues found in 369 source files
pytest -m "not stress" -n auto	6535 passed, 84 skipped, 1 xfailed — identical to baseline

Risks / notes for the reviewer

• No new tests (suite unchanged at 6535). This is the weakest of the eight on that axis: it is a
  rejection gate with no test asserting either that the noise is rejected or that ordinary prose is not.
  Recommend requiring a test before merge — happy to add one.
• Hardcoded list. The markers are Claude-Code-harness-specific (<system-reminder>,
  task-notification, tool-use-id). Reasonable for this project, but a reviewer may prefer it
  config-driven (write_gate.noise_patterns) so other harnesses can extend it without a code change.
• A memory that legitimately quotes one of these tags (e.g. a note about debugging the hooks) would be
  rejected. Rare, and the rejection is explicit with the reason in the message, not silent.
• The commit was reformatted with ruff format while preparing this branch (quote-style normalisation in
  _NOISE_TAG_RE); author preserved.