[guijacquemet]

External audit batch 34b, checked against main at bf733ee. I did not touch the image-type gate or block confirm (draft_answers.py, notebook_channels_grid.py, channel_units.py), and I did not touch the P21/P29 preflight wiring.

Each regression test below fails on bf733ee and passes on this branch. To confirm that, I reverted only src/ and ran the 7 new tests: all 7 failed, each on its intended assertion.

BLOCKER 3: one-click axes could confirm a preview that was never shown. STILL HOLDS

Reproduction: the header is readable but read_plane raises PlaneReadError. The preview is cleared, but quick_confirm stays enabled. _confirm_quick_axes() then returns True and sets preview_checked=True.

Fix:

• _render_preview records _preview_drawn: the mapping of the plane now shown, or None when no plane is shown (no image fits, or the read failed).
• The quick card enables Confirm these axes only when a PNG is shown and was drawn with exactly the offered roles. Otherwise it shows the read failure plus "Confirm these axes is not offered: no preview of these roles is shown."
• _confirm_quick_axes refuses the click, and never sets preview_checked, when the preview is empty, failed or stale.

Tests (tests/test_notebook_axes_quick.py):

• test_a_preview_that_cannot_be_read_is_not_confirmed_by_one_click
• test_a_preview_that_fails_later_withdraws_the_click

BLOCKER 4: an unchanged-folder fold could hide a changed-header refusal. STILL HOLDS

Reproduction (real workflow): publish DNA@C0 / ACTIN@C1, then rewrite ctrl/rep1/a.tif with the header names swapped, keeping the same file, shape and count. On reopen:

• PublishedState.changed is False.
• No unit has a reason.
• The screen folds under "Saved on …, nothing changed".
• The run's checks still give R_HEADER_CONTRADICTION.

Fix (root cause):

• observe_folders(..., metadata=) builds each unit's header evidence from the headers the preflight already read: the channel count and raw header names (physical_check.layout_of, without roles). A count the header leaves open stays unknown. A unit whose headers disagree with each other is marked ObservedUnit.uniform=False, which counts as channels_changed.
• published_state (notebook path) passes pre.metadata.files.
• PublishedState.changed now also covers rec.reopened, so every reason that reopens a repeat counts as a change.
• The unit is named in orange ("headers changed: ctrl/rep1 …"), and the step-1 chip counts folders with changed headers.

Second layer: when the run's checks have run on the published file (roles confirmed) and refuse it:

• the screen unfolds;
• "nothing changed" and "The run (step 12) can start" are not shown.

This also catches a refusal the reconciliation cannot see, such as a mask type added beside unchanged folders.

Tests:

• tests/test_notebook_row_fold.py::test_a_header_rewritten_under_the_same_files_keeps_its_row_open: real reopen, no injected reasons. The row stays open, R_HEADER_CONTRADICTION is visible, and nothing says the run can start.
• tests/test_notebook_row_fold.py::test_a_refusal_of_the_run_s_checks_unfolds_an_unchanged_reopen
• tests/test_reconcile_observe.py::test_the_headers_read_give_each_unit_its_channel_evidence

FILE 3: saved axis roles were offered after a header-only change. STILL HOLDS

Reproduction: after a run, a header replacement with the same image count still offered "the roles the earlier run of this experiment recorded".

Fix: with BLOCKER 4's evidence, state.changed is True, so _saved_axes() returns None. The card offers only the roles the headers name now, with "Some image headers changed since the experiment was saved … Review these afresh." Nothing is confirmed without the click.

Test: tests/test_notebook_axes_quick.py::test_a_header_change_does_not_offer_the_earlier_run_s_roles

FILE 1: there was no NaN-containing crop test. STILL HOLDS, and it found real defects

Defects found:

• Magenta non-finite pixels counted as "detail", so a NaN block outside the busy square took the crop: (0,0) instead of (256,640).
• A non-finite pixel only in the native crop was not named anywhere.
• The thumbnail note said "N non-finite pixels shown in magenta" twice.

Fix:

• busiest_square/_where measure finite thumbnail pixels only.
• crop_text names the crop's own non-finite pixels.
• display_panel names them once.
• A colour crop with non-finite values keeps its thumbnail and says why there is no detail.

Source arrays are never changed.

Test: tests/test_notebook_thumbnails.py::test_non_finite_pixels_inside_and_outside_the_crop_are_flagged_and_do_not_move_it

Coordinator item: bind the axis preview to the checked header. Fixed (independent of #254)

The axis-step preview read planes as the file is now, not as the folder check read its header.

Fix:

• The notebook folder check now calls preflight(..., bind_headers=True). The line and comment are identical to Audit batches 33, 34a, 34b: publish only what was answered, of the image checked #254's, so the two branches merge cleanly; I checked by test-merging origin/claude/batch33-fixes, and the notebook, reconcile and batch-33 tests then gave 420 passed.
• _render_preview reads a plane only while the file has the fingerprint the check took with its header, checked before and after the read.
• An image that is unbound or changed gives "… changed since the folder check read it: check the folder again." That counts as a failed preview: no image is shown, Confirm these axes is neither offered nor accepted, and preview_checked stays False.
• This branch does not merge claude/batch33-fixes and does not use its new helpers (check_unchanged/checked_fingerprint). The check is done locally with Fingerprint.of.
• tests/test_notebook_masks.py gets the same bind_headers=True one-line change as Audit batches 33, 34a, 34b: publish only what was answered, of the image checked #254.

Test: tests/test_notebook_axes_quick.py::test_an_image_changed_since_the_folder_check_is_not_previewed_or_confirmed

Gate

• python tools/check.py:
  • ruff check: clean
  • ruff format: 409 files already formatted
  • pytest: 5022 passed, 4 skipped (case-insensitive file system ×3, root file modes ×1)
  • code rules: pass, baseline not grown
• tests/test_spec_*: 1649 passed, 0 skipped

https://claude.ai/code/session_01G3qp9AxjpGPw9xfWAUJNqc

---

Generated by Claude Code