[guijacquemet]

What

For a run whose output folder holds a published morphobricks.toml, the experiment-wide v0.4 checks channel_names_differ and different_channel_counts no longer apply (run/description_run.py, without_global_channel_checks). The per-unit checks already in physical_check.py decide instead (R_HEADER_CONTRADICTION, R_CHANNEL_COUNT_MISMATCH, R_UNIT_NOT_HOMOGENEOUS; units with different channel lists are still R_HETEROGENEOUS_LAYOUT_NOT_ENABLED).

Refusal text now names the unit folders, the image, the image's header names (as read, quoted) and the channel's confirmed header names; the count mismatch also lists the declared channel IDs.

Runs without a description (v0.4 path) are unchanged: the global checks still apply. If the description cannot be read, the global checks are kept too.

Why

experiment-v1 section 4.3 and 8: header evidence is compared per unit and per C index, after NFC of both sides, with no trimming, case folding or similarity. A described experiment whose units spell a channel differently through confirmed aliases (fixture v11: DAPI / Hoechst 33342) was refused by the global check.

Regression evidence

New tests/test_spec_run_aliases.py (15 tests, added to EXPECTED_MODULES).

• On current main code (src changes stashed): 4 failed, 11 passed. The 4 failures are the alias run being refused (channel_names_differ), the undeclared-name refusal carrying the global code and lacking the unit/names text, the by-C-index refusal text, and the wrong-count unit also raising different_channel_counts. The 11 passes are the negative cases that must keep refusing, and the undescribed run.
• With the change: 15 passed.

Checks

python tools/check.py: 4241 passed, 4 skipped (skips not individually inspected; none are Spec modules, since a skipped Spec test fails the gate), code rules pass. Includes tests/test_parity_v0_4_0.py.

Needs an owner decision

1. Images with no header names: the Spec says they "need no evidence" (section 4.3, fixture v12), so they pass the header check and are not listed as unchecked. The request suggested listing them as unchecked; implemented the Spec text instead. Should such images be surfaced (for example a warning or a manifest list) even though the Spec does not require it?
2. An image with names at some C indices and none at others: each non-empty name is checked, empty ones need none. The Spec says "a non-empty header channel name", so this follows it; confirm that partial naming is intended to pass.

CHANGELOG: one entry added under Unreleased (keep both entries when merging main).

---

Generated by Claude Code