[guijacquemet]

What changes

rc.1, rc.2 and main all reported __version__ 0.2.0rc1 and the same features, so hosts could not tell them apart. This PR makes this build recognisable, for the pre-release v0.3.0-alpha.1 (tag and release are created by the maintainer after merge; nothing here touches tags or releases).

1. _version.py -> 0.3.0a1 (pyproject, schema tools_version, doctor read it from there; nothing else hard-codes it).
2. New feature id shm_order_v1 (shared-memory descriptor order honoured for C/F, anything else unreadable_image; behaviour from Worker: honour or refuse shared-memory order; memory-map big TIFF inputs #125, which already exists - PROTOCOL.md says so explicitly, this is the one id added after the change). No id for the TIFF memory map; PROTOCOL.md states "Tools >= 0.3.0a1 maps big TIFF files".
3. PROTOCOL.md: retention_prune is in rc.1 (the "not in 0.2.0rc1" claim is removed); runs.build_record/record_json/done_status arrive in 0.3.0a1 (also in rc.2 code, which still reports 0.2.0rc1); hosts check min_tools_version 0.3.0a1 or hasattr.
4. OPERATIONS.md: pre-release tag may be pinned for a human-test round; name rule (-rc.N <-> rcN, -alpha.N <-> aN); rc.2 did not bump and there is no rc.3; Windows file-lock caveat of the TIFF map; install/pin snippets say alpha.1. README gets the pre-release install line. CI compat job tag selection uses git -c versionsort.suffix=-alpha -c versionsort.suffix=-rc tag --list 'v*' --sort=-v:refname; checked in a scratch repo with tags v0.2.0-rc.1, v0.2.0-rc.2, v0.3.0-alpha.1: plain sort picks alpha.1; with v0.3.0 added the plain sort still picks v0.3.0-alpha.1 (the bug), the fixed sort picks v0.3.0, and alpha.1 beats rc.2.
5. conformance.json compat.version_order (+ HOST_PARITY.md copy) gets 0.3.0a1, 0.3.0. New frozen fixture tests/fixtures/schemas/tools_0.3.schema.json (0.3.0a1, three features); the compat matrix reads it (new host accepts it, old host reads it, a host needing 0.3.0a1 refuses the 0.2 fixture). Existing vectors and fixtures unchanged.
6. Release sandbox: playground_extras.py no longer hard-codes 0.2.0rc1 / v0.2.0-rc.1; --tools-pin argument (workflow passes its tools_pin input), default read from the installed app's post_install LC_TOOLS_SPEC line, and the version must equal the PEP 440 form of the pin. The workflow stays dispatch-only; tools_pin default is now v0.3.0-alpha.1.
7. docs/RELEASE_NOTES_0.3.0-alpha.1.md: draft notes to paste into the GitHub pre-release.

Tests changed (superseded by this change; none loosened or skipped)

• tests/test_run_record_builder.py: the assertion "not in 0.2.0rc1" is now "arrive in 0.3.0a1" (plus assertNotIn of the old wording); the expected FEATURES list gains shm_order_v1.
• tests/test_compat_matrix.py: SHA256 entry for the new fixture; matrix prints 14 ok lines instead of 11 (3 new checks); new assertions on the 0.3 fixture.
• tests/test_versions.py: needle v0.3.0-alpha.1 added to the OPERATIONS.md check.
• tests/test_tools_features.py: new test for the id list and the PROTOCOL.md statements.
• tests/compat_matrix.py, tests/release_sandbox/playground_extras.py: as described above (black reformatting only otherwise).

Not verified

• Local run (Linux, Python 3.11, running as root, private venv): 637 tests, 1 failure, test_hardening.RegistryNamesAndPriority.test_foreign_owned_or_group_writable_shared_entry_is_skipped, which depends on file ownership and cannot pass as root; it passes on the CI runners. black (26.10.0 as pinned), ruff, mypy, vulture clean (local ruff is 0.15.20, not the pinned 0.16.10; CI lint is green).
• CI on head 253617b: full matrix (workflow_dispatch) https://github.com/CellMigrationLab/LabConstrictor-Tools/actions/runs/38417011793 and the PR's reduced matrix https://github.com/CellMigrationLab/LabConstrictor-Tools/actions/runs/38417004716: lint, compat and every test job green.
• The CI compat job ran against v0.2.0-rc.2 (newest tag today); the sorting fix is verified in a scratch repo only, not in a repo that has v0.3.0 yet.
• The new playground_extras.py basic step and the pin reader in release_sandbox.yml were not run (the workflow is dispatch-only and needs a released app; the pin regex and tag-to-PEP 440 mapping are untested on a real installer hook).
• The release notes are a draft; wording of "Worker exit code 3 is called an import failure only when the worker said so on stderr #124" follows the task description, not a re-read of the PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y2SN8SuEakTdiWy3F2WUqN

---

Generated by Claude Code