The landing scan for tune-bench-comparison: five gates, and nine tenths of it is run output

[syncytium2]

Tony asked on 2026-09-19 what has to be true before the weekend's branch can reach main. Every figure is measured, not estimated. Re-measured at 17:40 EDT after gate 1 was repaired and #660 merged the architecture figures in; the first column is the original scan against main at ab4eef1.

The branch is not dirty

It merges with no conflict, sapper --all reports 0 blocks (78 warnings, nearly all SAP015's "data is plural" on prose), check_quotes is clear.

And the insertion count is misleading in a way worth correcting before anyone reacts to it:

	first scan	now
commits ahead of main	59	66
files	454	475
insertions	54,264	56,382
code, tests and tools	21 files, 3,742	25 files, 4,162
run output under docs/learned/	429 files, 49,509	440 files, 51,120

Nine tenths of the diff is what the runs produced.

Five gates, one of which is now closed

1. The base is red. Repaired. WSMIP064 marked the module serial in b9752a9 and the base went green on run 1624. The fair comparison's four architectures, drawn through draughtsman at one slot #660 then passed on the repaired base and merged, which is why the counts above grew.

   Half of what this gate prescribed was wrong, and that is recorded rather than quietly dropped. It also said to drop the inner --jobs to 1. 064 measured first: the quick run takes 276 s at four jobs and 559 s at one, and 559 against a 600 s budget leaves no room on a slower runner. Jobs stayed at four; only the mark changed. CI came in at 419–545 s of the 600.

   A consequence that arrives on main with this branch: a serial module runs after the parallel pass with the runner to itself, so any branch carrying this fixture now runs a ~2-hour suite rather than ~10 minutes — Goal 2: nested-CV tuning of the nets against every-knob coded detectors, on the GPU and the bench #642 took 2 h 06, The fair comparison's four architectures, drawn through draughtsman at one slot #660 took 1 h 47. Afterwards that applies to every PR in the repository. Whether the fixture is trimmed first is outside this scan, but it is worth asking before the landing rather than after.

2. Landing this branch also lands Every learned model against CoactDetect; gauge and chorus evaluated, chorus repaired #596 — it is an ancestor, 83 files and 24,820 insertions of the total. Goal 2: nested-CV tuning of the nets against every-knob coded detectors, on the GPU and the bench #642 says its merge is Tony's call because it registers the learned models in the lab server and the model picker, and Every learned model against CoactDetect; gauge and chorus evaluated, chorus repaired #596 itself asks to be held if they should stay out until a real-recording check. Merging here answers that by accident.

3. One personal path in a public repo, in HANDOFF-workstation-tuning.md, which SAP004 cannot see because it matches forward slashes only (The personal-path guard reads one slash, and a Windows run directory is the other #659's todo). Still open — the username is still at line 682. One line, and it must be fixed before landing: a push here is publication.

4. The root handoffs need the three-way decision docs/handoffs/README.md sets out. Still open — five handoff files sit at the root.

5. Ask what the landing is for. The seven stranded architectures are 7 files of the 475. If un-stranding them is the goal, a nets-only branch does it today and inherits neither the reserved ruling nor the run output. Landing everything is a larger, separate act — both are reasonable, they are just not the same one.

Docs only, one file. Sapper, check_quotes and check_pipelines clear.

🤖 Generated with Claude Code

https://claude.ai/code/session_017hQX6iESxBsQk7Jan875e3