fix(cues): session-2 precision pass — memory-test rule, word-boundary clips, 0.35 dedup backstop

[xerhab]

What

Fixes the five cue-quality failure classes surfaced by the second real recording (21 min of entity-dense YouTube tech audio, conv 7ca38157, now a harness fixture alongside the family session and the frozen 12):

1. Cross-generation confabulation — a model number newer than the weights got a sibling's specs pasted under its name ("Galaxy Z Fold 8 … launched August 2023"). Three prompt formulations iterated (v10→v12); the shipped v12 reframes the rule as a specific-memory test the model can actually run ("family resemblance is not knowledge — no specific memory, no cue"), extended to vaguely-recognized names/acronyms, plus an evidence-side rule (an article about the predecessor covers nothing about the successor) and a stale-date rule (never correct anniversaries/ages from the internal clock).
2. False corrections — the correction trigger now requires positively knowing the truth, and never cues that a name the speakers used "does not exist" (production had inverted truth on "Jet Grind Radio", a real US regional title).
3. Paraphrase duplicates — avoid-list now bans re-approaching a surfaced subject from a different angle; session substance-dedup backstop lowered 0.5 → 0.35, calibrated on 90 real production cues (retitled rewords measure 0.30–0.38; closest genuinely-distinct pairs 0.26–0.27 — 0.35 is the safe floor).
4. Mid-word truncation — 8 of 51 production bodies were chopped by the hard body[:240] slice; replaced with a word-boundary clip + ellipsis and an "under 200 characters" nudge (final replay: 2/401 clipped, cleanly).
5. Stale-topic cues — newest-turns recency rule (fast topic-switching audio kept getting cues for topics left 60+s earlier). Also: lifestyle-advice ban, bilingual mishear rule (a stray foreign-looking word is the speakers' other language, not Dutch).

Verification

Replayed the full 14-conversation export (985 attempts) through the deployed gpt-oss-120b (medium effort) per iteration, hand-bucketed both real sessions each round, self-judged comparatively — full record in scripts/cue_eval/RESULTS-2026-07.md:

run	cues	novelty	relevance	accuracy	judge dups	garbled ctrl	question conv
v9 baseline	461	1.87	1.93	1.92	28	10	1
v12 (this PR)	401	1.88	1.90	1.94	17	5	5

Hand-bucketed on session 2: wrong/confabulated ~15 → ~5±2 per run, paraphrase-dup pairs 6+ → ~1-2, stale/dictionary near zero, and the direct-question conversation went from near-mute (1 cue) to properly answered (5). Volume lands at 87% of baseline with the cuts concentrated in the failure classes.

Known residual (documented): first-party sibling-spec transfer (the Z Fold 8 case) survives every prompt formulation — the model genuinely believes it knows the product. The fix is the already-planned retrieval cross-check of entity cues, not more prompt text.

Tests: full API suite 306 passed, coverage 96.04% (gate 85%), ruff clean. New tests pin every new prompt rule, three truncation cases, and a session-level regression test using the measured 0.38 paraphrase pair.