[guijacquemet]

Fixes the items of the rules audit (/home/user/AUDIT_rules_tools_template.md) that had to land before the tag v0.3.0-alpha.1. Smallest changes, no refactoring. Two commits: code, docs and tests; then CI hardening. Not merged.

What changes, per item

1. Silent fallbacks

• convert._memmap_tiff: every "read it the plain way" return goes through _not_mapped(path, reason). For a file of 8 MiB or more (MEMMAP_MIN_BYTES) it writes one warning, TIFF <name> was read into memory, not memory-mapped: <reason>, to the log file (log.warning) and to the worker's stderr (LabConstrictor worker: ..., like the worker's other notes). Reasons: compressed, tiled, multi-series (n series), big-endian byte order, truncated, not contiguous, pyramid levels, OME/other series kind, pixel type, LC_TIFF_MEMMAP=0, or a failed mapping (exception type and text). Files under 8 MiB, and mapped files, say nothing. The refusal conditions are the same as before, only split so each has its reason (no behaviour change for what is mapped).
• docs/OPERATIONS.md "Fallbacks (intended)": rows for the TIFF fallback and for Worker._orphan_exit.
• worker._orphan_exit: kept ignore_errors=True (a failing remove must not stop the worker leaving), but a folder that still exists is logged with log.warning to the log file only (stderr has no reader once the host is gone; _complain there could raise on a closed pipe and stop the exit).
• Rechecked the 17 handlers of the audit: convert.py:181 (OverflowError, refused two lines later), the two ProcessLookupError, registry.py:417 (race, commented, entry reported elsewhere), retention OSError returns (end in the report skipped or the documented lock exit code), cli FileExistsError, prune guard, diagnostics, exporter, protocol, conformance: all are told elsewhere or are correct control flow and are left as they are. Only the two above were silent without a reason.

2. Ledger and HOST_FEATURES

• REGRESSION_LEDGER.md: rows for Worker exit code 3 is called an import failure only when the worker said so on stderr #124 and Worker: honour or refuse shared-memory order; memory-map big TIFF inputs #125 with their guards; a row for the TIFF fallback (Audit fixes before v0.3.0-alpha.1: told TIFF fallback, ledger, docs, test hygiene, CI hardening #129); all 26 "this PR" replaced by real PR links (looked up with the PR search: Fix: error codes and validation (F1 F5 F6 F11 F12 F15 F16 F20) #41, Fix: worker and client lifecycle (F13 F17 F18 F19 F21) #42, Fix: labels_to_shapes never drops a label that simplification collapses (F14) #43, Fix: Windows quoting follows the C runtime backslash rules, shared quote vectors (F10) #44, Fix: table inputs read floats exactly, keep names as written, NA words stay text, readable errors (F2 F3 F4 F7) #45, Fix: notebook form has no silent number clamp, and optional parameters can be left unset (F8 F9) #46, Fix: Windows quoting (F10) and table inputs (F2 F3 F4 F7), merged and conflict-resolved together #47, Tools versioning: 0.2.0rc1, tools_version/features, reserved-input tolerance, compat check, compat CI job #110, Conformance: conformance.json (packaged), validating loader, host loaders documented (phase 0.3, PR B) #113, Notebook form: refuse an integer beyond 2^53-1 with the sentence of spec 6.2 (NUM-02, NUM-03) #116, Fix main CI on Windows/macOS Python 3.10 (wheel test, conformance CLI test) #118, Record two decisions of 13 Oct: changed-size sentence, crash after Cancel (spec, data, fixtures) #121, runs: build_record, record_json, done_status for hosts that write their own run record #122; F2/F3/F4/F7 and F10 carry both the fix PR and Fix: Windows quoting (F10) and table inputs (F2 F3 F4 F7), merged and conflict-resolved together #47, which merged them); the "fix PRs pending" section is gone, its 8 rows moved into the table, and a note says tests/roundtrip_known_failures.py is empty (FINDINGS = {}, KNOWN = {}).
• HOST_FEATURES.md: rows for the shared-memory order (shm_order_v1) and the TIFF memory map. The host columns say "same" (worker-side behaviour); whether each host sends F-ordered blocks is not checked here.

3. Install from the tag. README.md (the main.zip block removed, the tag block kept), OPERATIONS.md Install (Tools by tag; the Napari plugin has no tag and stays refs/heads/main, labelled as development version), LABCONSTRICTOR_INTEGRATION.md line 15, HUMAN_TEST_PROTOCOL.md line 207 (the URL only; the napari[pyqt5] text there is untouched, as asked).

4. OPERATIONS contradictions

• Line about temp folders: hosts never prune, they write the markers and call the Tools CLI.
• features.py docstring: states the exception (a flag added later for already shipped behaviour is allowed when it is the only way to tell versions apart; shm_order_v1).
• Tags: lightweight tags created by the GitHub release are the standard; procedure steps 3 and 4 merged ("create the release with the new tag name on the merge commit"); "Never move or recreate a tag" kept.
• Windows: registry trust checks (owner, group/other-writable, interpreter inside the prefix, world-writable folders) are POSIX only; "Windows is unchecked" and why.
• Playground rule written in TESTING.md ("The Playground rule", with where it is checked) and pointed to from OPERATIONS.md.

5. Test hygiene. tests/README.md no longer hard-codes a count (and names the permanent skips); the dead try/except ImportError and skipIf(retention is None) in test_conformance.py removed (retention is imported directly, the cli/runs fallback branch too); test_regressions_run1_gc now asserts gc.garbage == []; TESTING.md lists smoke_migration. Also in TESTING.md: the sabotage claim now says "in the table" and names what is covered, and the known-failures row says the list is empty.

6. CI hardening (second commit). The only third-party action is conda-incubator/setup-miniconda (smoke_install, smoke_migration): pinned to fc2d68f6413eb2d87b895e92f8584b5b94a10167 # v3.3.0 (SHA of the tag v3.3.0, equal to the old v3 tag, resolved with git ls-remote and fetched to confirm the commit exists). tj-actions/changed-files is not used in this repository (it is in the template). GitHub's own actions/* stay on major tags; TESTING.md says so. release_sandbox: new inputs fiji_ref and qupath_ref (default main, passed to git clone --branch through env; the default branch of both repositories is main, so behaviour is unchanged); napari_ref description says what it is. TESTING.md: smoke workflows are dispatch-only (no schedule, no PR trigger; smoke also on pushes to claude/smoke-bridge) and a table of which ones default to branch names.

Tests changed

• Added tests/test_input_loading.py::MappedTiff: test_a_file_that_could_not_be_mapped_says_why_in_the_log_and_on_stderr (compressed, tiled, multi-series, big-endian, not contiguous, truncated; one line, file name, reason, same text on stderr), test_the_environment_switch_is_told_for_a_big_file, test_nothing_is_said_for_a_small_file_or_a_mapped_one.
• Added tests/test_worker_orphan_exit.py (2 tests).
• Added 3 rows to tests/test_sabotage.py (log line removed, small-file gate removed, orphan log removed) and updated the existing row "pages of a TIFF stack are not checked for gaps" because the line it breaks moved. Sabotage-checked: each new row's break makes its slice fail (run with LC_SABOTAGE_ONLY); while writing the compressed-file test it also caught a wrong reason ("no contiguous pixel data" for a compressed file), fixed before the commit.
• Changed tests/test_conformance.py (guard removed, nothing else), tests/test_regressions_run1_gc.py (assertion added). No test was skipped, loosened or deleted.

Not verified

• Windows and macOS behaviour of the new tests beyond what CI shows (green on all jobs); the stderr line goes through the real worker only in the unit tests, not in a host window.
• Whether Napari, Fiji and QuPath show the new stderr line in their Details windows (they show the worker's last stderr lines; not run here).
• The release_sandbox change is not exercised (dispatch-only, needs a released app); the YAML parses. The pinned setup-miniconda SHA runs only in smoke_install / smoke_migration, which were not dispatched.
• The PR numbers in the ledger come from titles and commit history; F2/F3/F4/F7 and F10 are credited to both the single PRs (Fix: table inputs read floats exactly, keep names as written, NA words stay text, readable errors (F2 F3 F4 F7) #45, Fix: Windows quoting follows the C runtime backslash rules, shared quote vectors (F10) #44) and Fix: Windows quoting (F10) and table inputs (F2 F3 F4 F7), merged and conflict-resolved together #47.
• Third-party-action review covered this repository only; the template's tj-actions/changed-files@v44 and the template pin are separate work.

CI

ci on the PR head and a full-matrix workflow_dispatch run (6 test jobs, lint, compat) are green: https://github.com/CellMigrationLab/LabConstrictor-Tools/actions/runs/38425033947

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y2SN8SuEakTdiWy3F2WUqN