fix(health): stop reporting time series as supersede candidates, and hot topics as sediment gaps

[MXAntian]

Both signals in memory-health were firing on the wrong thing. On a 9301-row library the supersede band listed 125 pairs and the repeat-query list 4; a hand review found ~1 real item across both.

A report that is 99% false positive doesn't get acted on — it gets skimmed, and the one real pair goes with the rest.

(a2) — time series read as rewrites

The band exists to catch "an already-replaced version stays active and can be recalled as if current." A time series has no replaced version: every entry is still true about its own moment, and superseding one destroys the sequence.

The old guard only excluded pairs where both rows were concrete_trace. Series written at semi_abstract sailed through — nightly metric snapshots, timestamped log lines from one evening, a request and its reply, a decision and its refinement three minutes later.

isLikelySeries now takes three signals, any one sufficient:

1. both concrete_trace (the old rule, kept)
2. written inside one working session (default 8h)
3. both summaries carry a temporal marker and the markers differ

Signal 3 requires the markers to differ on purpose: two rows citing the same date are one event described twice, which is exactly the rewrite we want to surface.

Calibrated against 10 hand-classified pairs pulled from the live library rather than from intuition — the widest same-session false positive sat 6.8h apart, the closest true rewrite 5 days. Those 10 shapes are now series-detection.test.mjs, so the next change to the predicate has to argue with the data.

Effect: 125 → 22 candidates. The one pair the band exists to find (千夏不爱吃香菜 vs 千夏讨厌香菜的味道, 5d apart) moved from buried to first.

(d) — frequency mistaken for evidence of a gap

Repeat queries were labelled sediment-worthy on frequency alone, while SUM(hit_count) was already selected and then ignored.

Frequency measures how hot a topic is, not whether it's answered — and hot-and-answered is the normal case, since hooks fire recall on every prompt containing the word.

Measured before fixing: every freq>=3 query in the 7d window averaged 5–20 hits. The signal was 100% false positive. watchdog ×12 was the loudest, and recall returns three good cards for it.

Now filtered to avg hits < 1, with the answered ones reported as a count so the information isn't lost. Live: 4 → 0 flagged, 4 hidden as hot topics.

Verification

• series-detection.test.mjs — 11 new cases, red before the change (no such export), green after
• memory-health.test.mjs — 68/68, unchanged
• both wired into CI

Risk

The 8h same-session window is calibrated on 10 samples. It will miss "genuinely rewrote the same fact within one day" — that costs one missed candidate, against 103 removed false ones. If a real instance shows up, tune seriesSameSessionHours rather than removing the rule.