The newest export was on disk for six days and no file here named it

[syncytium2]

Groundwork for running the full senktide and TTX cohorts through the loop. The run itself is still going; this is everything it needed that did not exist, landed on its own so the branch does not carry code and results together.

The finding that started it

Tony asked for the full senktide and TTX data set "with excluded artifacts", and the first lookup turned up an export the repo had never heard of. interface2 shipped 2026-09-03_revised_2v_long_STEPS_EXCLUDED six days ago — artifacts removed, pre-split into SENKTIDE (29 recordings) and TTX (38) — and its own README says "for any new analysis, use this folder." No file here named it. current_export.toml, which exists because a session once could not tell which folder was current, still pointed at the 2026-08-18 export.

It is declared now as three roles beside default, not instead of it. Moving default would silently re-point every existing analysis and every parity fixture at a different corpus; that is a decision for Tony, and the INDEX row says so.

Three things the run could not honour

	was	now
K = 10% floored at 3	not expressible — k_from_fraction clamps at 1 and leaves the clamp to the caller, and nobody had written the caller	--k-floor on assess, assess_archive.py and derive_spec.py; every result records whether the floor bound there
a spec from a percentage	derive_spec selects one absolute-K column, which drops every recording whose K landed elsewhere	--k-percent aggregates across recordings at each one's own count
group facets	one page, all detectors stacked, no groups	--per-detector --facet group

The floor is not cosmetic here. The APV+CNQX+GZ pilot filed this same gap and got away with it because that cohort's field sizes made 10% and floor-3 the same K on every recording. On this corpus the floor binds on 26 of 84 recordings, so the population is governed by two K rules at once — which is why a bound floor is recorded per recording rather than absorbed into the number.

And selecting a column would have been quiet. 10% floored at 3 resolves to K=3 on 55 recordings, 4 on 15, 5 on 12 and 6 on 2. The median column alone holds 55 of 84; the other 29 would have left the spec with nothing said about it.

k_source reads percentage_unreviewed, never mahice. MAHICE was skipped on instruction, so this K has the shape of one a person would set and no person set it, and a consumer has to be able to tell those apart.

Two silent defects in code that was already shipping

The shared PNG renderer clipped at a literal 1120.0 — the old viewport as a constant — so widening the viewport widened the window and clipped back anyway. A four-facet page rendered as three facets and a sliced fourth, with no error and a file that looks finished unless somebody opens it. Same shape as the height bug documented two comments above it in the same function.

make_detector_table.py reads rates() by tuple position across a module boundary. Adding the group field turned it into a TypeError — the good outcome; a shifted numeric field would have been wrong numbers and no complaint. Fixed, with the reason written at the call site.

Checks

Full suite green — 2270 passed, 46 skipped. Sapper clear, check_quotes clear. New tests pin the floor (that it applies after the fraction meets the recording, that a bound floor is recorded rather than absorbed, that it is refused under an absolute K) and the facets (group carried unchanged, unlabelled facet drawn last, inks stable across a run's pages).

🤖 Generated with Claude Code