Why a third of chorus_norm's tuning fits do not train: the head is never woken at lr 0.03 (landed with murderboard flags)

[syncytium2]

In both draws of goal 2, about a third of chorus_norm's tuning fits made one call per recording. This PR adds the diagnosis of why: a readout page, the tool that builds it, and the data the page is rebuilt from.

Landed with open murderboard flags, at Tony's call. Two review rounds ran, each with all 11 roles. Blocking findings went 3 → 4, so the process stopped and escalated. The second round's repairs are applied but have not been blind-reviewed.

What it finds

• At lr 0.03, chorus_norm's head is never woken.
  • Every replayed fit starts with its 8-layer head mostly silent: 5 to 8 of its layers pass nothing that varies.
  • The working fit's head wakes by step 50. The replayed collapsed fit's never does.
  • Measured away from the recording's ends, every collapsed inner fit has a silent head layer (153 of 153 chorus_norm, 75 of 75 chorus_gain_norm) and no working fit does.
• Rates. Across both draws, 292 of 396 inner fits collapse at lr 0.03, against 7 of 468 below it.
• Mostly not PR Every learned model against CoactDetect; gauge and chorus evaluated, chorus repaired #596's failure. By Every learned model against CoactDetect; gauge and chorus evaluated, chorus repaired #596's own test, how far one onset moves a vote, most collapsed chorus_norm fits' votes still respond: median 0.40, against 0.0003 for plain chorus. 4 of 153 are as deaf as plain chorus, and 19 of chorus_gain_norm's 75.
• The same collapsed fit, replayed exactly, trains at lr 0.003 or 0.01, or at 0.03 with a 200-step linear warm-up. That warm-up brought training loss below 0.5 in 6 of 7 other hand-picked collapsed runs.
• Tuning. It picked no lr-0.03 configuration for chorus_norm under either selection rule. For chorus_gain_norm, the selection on F1 alone picked lr 0.03 in 6 of 8 folds, and both of its collapsed refits came from those picks.

What is in it

• tools/diagnose_chorus_collapse.py, with subcommands table, census, trace, replay and page. replay re-runs a fit bit-exactly against its checkpoint and can change one thing at a time.
• docs/learned/chorus_collapse/: the page, plus the JSON it is rebuilt from.
• tests/test_diagnose_chorus_collapse.py: rebuilds the page byte for byte, checks figure numbering, and checks for personal paths.
• The todo docs/todo/2026-09-19-chorus-norm-does-not-train-at-lr-0.03.md, which holds the repair decision.
• An INDEX row and glossary entries.
• The murderboard record docs/reviews/chorus-collapse_2026-09-19.md and both rounds' role reports.
• A merge of main, so the fair-comparison report the page links to is present.

Open flags

• ⚠ The round-2 repairs are not blind-verified. These cover a rebuilt Figure 1, a new Figure 4, a merged Figure 5, the census and replays measured away from the recording's ends, and rewritten sections 1–6.
• ⚠ The replicate report's reading is wrong. It reads the 111 fits that collapsed in both draws as the same fits failing twice, but the configurations' rates alone predict that many. This page corrects the reading. The report's own fix is in the todo.
• ⚠ Warm-up rescues are judged by training loss, not by calls on held-out recordings.
• ⚠ "Silent" uses an absolute threshold. Some of the silence at the starting weights may be scale.
• ⚠ Role 7's refactors were declined. These were running replays through train() with hooks and sharing the collapse rule with the replicate report's builder.

docs/goals/ is untouched (#664 owns it).

🤖 Generated with Claude Code