fix(code-review): make override-cache tests independent of wall-clock date

[wongk]

Problem

Three tests in test_code_review_helpers.py were failing on main with no accompanying code change:

FAILED TestOverrideCache::test_verify_prepare_short_circuits_on_valid_override
FAILED TestPR114ReviewFixes::test_prepare_then_consolidate_routes_override_to_verified
FAILED TestPR114ReviewFixes::test_prepare_then_consolidate_writes_re_asserted_to_stats

They are time bombs. Each writes an override with a hardcoded asserted_at of 2026-05-29T22:00:00+00:00 and then asserts the override is honored — landing in override_hits, or routed to a RE_ASSERTED verdict.

The overrides cache has a 90-day TTL (CACHE_TTL_DAYS["overrides"]), enforced by _override_is_expired. The tests passed while that fixed date was recent. Once wall-clock passed 2026-08-27, _override_is_valid correctly began returning False, override_hits came back empty, and the assertions broke.

The production code is correct — expiring a 100-day-old override is exactly the intended behavior. Only the fixtures were wrong.

Fix

Derive asserted_at relative to now, so the tests exercise the honored-override path whenever they run.

The file already had a helper for this — _stale_cached_at(days_ago), returning an ISO timestamp N days in the past — and two sibling TTL tests already used the relative idiom inline. Rather than add a near-duplicate, this reuses the existing helper and renames it to _iso_days_ago() so the name matches its already-generic docstring (it is used for both within-TTL and past-TTL fixtures, so "stale" was misleading). All 6 existing call sites updated, plus the two override TTL boundary tests, which now delegate instead of inlining datetime.now(timezone.utc) - timedelta(...).

Scope check: is this the only time bomb?

Swept every hardcoded timestamp feeding TTL-checked fields across all plugins' tests. Two remain, both safe by construction:

Location	Value	Why it can't rot
test_code_review_helpers.py:3560,3586	cached_at: 2025-01-01	Deliberately ancient; tests assert overwrite/retention, which an older date only reinforces
test_code_review_helpers.py:10071	asserted_at: 2026-05-29	Write/load round-trip equality; never passes through TTL logic

The rot-prone pattern is specifically "hardcoded recent timestamp expected to be within TTL", and all three instances are fixed here.

Also included

/update-documentation is mandatory before push in this repo, and its README-verification step flagged two pre-existing inaccuracies in plugins/code-review/README.md. Both are unrelated to the test fix but verified correct, so they are included rather than left wrong:

• Golden-fixture count said "3 full scenarios + 6 README-stubs"; on disk it is 4 full + 3 stubs.
• Four files present on disk were missing from the architecture tree: config/cli.json, config/stages.json, signal_taxonomy.json, conftest.py.

Verification

• pytest plugins/ — 2107 passed, 3 skipped (was 3 failed, 2107 passed)
• uv run ruff check . — clean
• uv run pyright — 0 errors
• Both TTL boundary controls still pass in the right directions: 120 days → invalid, 30 days → valid

No production code changed — test fixtures, one test-helper rename, README, CHANGELOG, and a PATCH version bump (3.7.0 → 3.7.1).

🤖 Generated with Claude Code

https://claude.ai/code/session_011QjUMmrqfGU4QNXDDRLPrN

[shafty023]

Beginning a code-review. Please allow time for me to finish before merging the PR