fix(cut-lens): record run 2 — the distribution gate HALVED coverage; stop tuning

The medoid + effect-size gate was meant to buy coverage after run 1 returned
"no separating orbit" for six of fourteen boundaries. It did the opposite:
boundaries with any voting orbit went 8/14 -> 4/14. Requiring cross-side
distance to beat within-side spread is a stricter test, not a better-aimed one.

The error table is seductive and is not evidence: u=0.85 scores median 0.4s,
worst 0.7s — but n=2, and those two boundaries (#8, #10) already had ear
answers. Four survivors cannot validate five candidate fractions, and a metric
computed on the cases that were already easy measures nothing.

The durable finding is the pattern across all three attempts: #13 Vague de
CRIME and #15 REVOLUTION have produced ZERO voting orbits in v2, v3 run 1 and
v3 run 2 — three gates, two estimators. That is not a tuning failure. Those
transitions have no orbit whose timbre differs materially across the switch;
the outgoing and incoming tracks share their sounds, so a timbral lens is
structurally blind there just as an activity lens is blind to a crossfade. A
fourth loosening can only admit noise, because there is no signal in this
feature space.

Declared a dead end in the docstring rather than tuned again. Two ways out are
recorded: change the feature (harmonic/key change, or sample-bank identity from
the score, which differs even when the spectrum does not), or accept #13/#15 as
ear-only calls — two timestamps, seconds of listening each.

Candidates that DID survive are kept in cut_candidates_v3run2.json for #8, #10,
#11, #12 as ear-confirmation inputs, not as boundaries.

armada/tide-table/cut_candidates_v3run2.json

+[
+  {
+    "track": 8,
+    "title": "Eh ouais je Funk",
+    "nominal": 2720.0,
+    "onset": 2675.75,
+    "takeover": 2726.25,
+    "mid": 2710.75,
+    "voters": 4,
+    "truth": 2726.4
+  },
+  {
+    "track": 10,
+    "title": "Gimme Acid",
+    "nominal": 3252.3,
+    "onset": 3224.8,
+    "takeover": 3268.3,
+    "mid": 3229.3,
+    "voters": 3,
+    "truth": 3269.0
+  },
+  {
+    "track": 11,
+    "title": "Ghosts in the Toilets",
+    "nominal": 3560.2,
+    "onset": 3515.7,
+    "takeover": 3557.2,
+    "mid": 3554.2,
+    "voters": 3,
+    "truth": null
+  },
+  {
+    "track": 12,
+    "title": "Mafia",
+    "nominal": 3681.6,
+    "onset": 3667.1,
+    "takeover": 3688.1,
+    "mid": 3680.1,
+    "voters": 3,
+    "truth": null
+  }
+]
\ No newline at end of file

armada/tide-table/cut_lens3.py

@@ -82,6 +82,35 @@ Usable output today: #11 clusters tight (3553.7-3557.2, all five fractions
 within 3.5 s => a SHORT crossfade, so the nominal 3560.2 is ~4 s late) and #12
 spreads wide (3662.8-3694.6 => a long fade, nominal 3681.6 sits mid-span).
 Both are candidates for ear confirmation, not answers.
+
+>>> RUN 2, same day — the distribution gate made coverage WORSE. STOP HERE. <<<
+
+Framing the references, taking medoids instead of means, and gating on an
+effect size (cross-side distance must beat within-side spread) was supposed to
+buy coverage. It did the opposite:
+
+    boundaries with any voter:  run 1  8/14   ->   run 2  4/14
+    #13 Vague de CRIME, #15 REVOLUTION: no voters in EITHER run, nor in v2
+
+The accuracy figures look spectacular and must not be believed: u=0.85 scores
+median 0.4 s, worst 0.7 s — on n=2, and those two are #8 and #10, both of which
+already had ear answers. Four surviving boundaries out of fourteen cannot
+validate five candidate fractions. A metric computed on the cases that were
+already easy is not evidence.
+
+THE REAL RESULT IS THE THREE-VERSION PATTERN, NOT ANY ERROR NUMBER.
+#13 and #15 have never produced a single voting orbit across v2, v3 run 1 and
+v3 run 2 — three different gates, two different estimators. That is no longer
+plausibly a tuning failure. It says those transitions have NO orbit whose
+timbre differs materially across the switch: the outgoing and incoming tracks
+share their sounds, so a timbral lens is structurally blind there, exactly as
+activity-based detectors are blind to a crossfade. Loosening the gate a fourth
+time can only admit noise, because there is no signal in this feature space.
+
+Do not tune this again. Either change the FEATURE (harmonic/key change, or
+per-sample onset identity from the score's sample banks, which do differ even
+when the spectrum does not), or accept that #13/#15 are ear-only calls. The
+cheap honest move is the latter: two timestamps, seconds of listening each.
 """
 from __future__ import annotations