[guijacquemet]

What

Roadmap phase 0.3 part B ("87a"): the cross-host rules as data, plus the loader.

• labconstrictor_tools/conformance.json, packaged with the wheel ([tool.setuptools.package-data]; a test builds the wheel and checks it contains the file).
• labconstrictor_tools/conformance.py: load() (validated, cached), validate() (hand-written schema, stdlib only, returns every problem), path(), and python -m labconstrictor_tools.conformance --path | --validate for hosts that are not Python.
• docs/CONFORMANCE.md: layout of the file, how each host's test loads it (Tools, Napari, Fiji via LC_CONFORMANCE, QuPath), the rules for a host test, and the gaps. HOST_PARITY.md section 2, PROTOCOL.md and the spec header now point to it.
• Contents, only what is already decided in SPEC_WRONG_SCIENCE (or carried by Tools code): limits (number limits 2^53-1 and 1e15, results_kept 20, run_records_kept 50, lookup timings; each with status: decided|default), host_internal_codes (the four new ones), messages (templates), compat vectors, and wrong_science: coordinate frame (rule, placement formula, CF-01..CF-14), calibration (unit table, micro-sign and damaged-sign rule, anisotropy, default-dpi rule, _calibration shape, CAL-01..CAL-18, CAL-T1..T9), retention (marker names; RET-01..09, 12, 13 as structured fixtures, HRET and RET-10/11/14 as host fixtures), run identity RUN-01..06, async ASY-01..08, number limits NUM-01..05.

Tests (tests/test_conformance.py)

• Schema: 28 kinds of damage are each named (missing key, wrong status, duplicate id, positional placeholder, unordered versions, ...); validate never raises on odd data; a malformed file makes load() raise listing all problems.
• Data against the code: limits vs notebook.MAX_FORM_INTEGER and retention.RESULTS_KEPT/RUN_RECORDS_KEPT; every compat vector through compat.check_schema; every message template is a sentence of the spec; host-internal codes match the spec text.
• Data against plain-Python oracles: world coordinates of CF-01..06 (scale/translate/affine/shear), the rectangle mask by an even-odd rasteriser, the CF-09 crop arithmetic, c = x0 + (x + 0.5) * d for every CF-14 variant, unit conversions (CAL-03, 05, 15, 16, 17, 18), the default-dpi tolerance.
• Retention fixtures RET-01..09, 12, 13 are built from the JSON and run against retention.prune and the real CLI (4 concurrent processes, lock held by another process, symlink).
• Sabotage: 9 new rows (conformance:*) in tests/test_sabotage.py, all caught. Ledger row added in a different place from PR A's.
• black, ruff, mypy, vulture clean; full suite under coverage run as a non-root user on the merge result (main merged in, after PR A was merged).

Gaps found in the spec (not invented; full list in docs/CONFORMANCE.md "Gaps")

1. HOST_PARITY sketch keys with no decision in SPEC_WRONG_SCIENCE are left out: max_shapes, table_rows_shown, region_max_objects, max_export_bytes, max_display_bytes, cancel_grace_s, worker_orphan_grace_s, worker_idle_s, log limits, error_codes, notice_codes, worker_exit_hints, number_text (the Tools CLI itself diverges, TOOLS-7), quoting/command_vectors (Windows trailing backslash is "TO DECIDE"; tests/quote_vectors.json stays the shared source), results_text, trust_vectors, host_error.
2. The axis order of rect is not stated (the mask shows it is [y, x], while protocol shapes are [x, y]); CF-09/CF-14 mix [width, height] and [rows, cols].
3. CF-07, 08, 10..13, RUN, ASY, HRET and RET-10/11/14 are prose sequences; there is no machine-executable vocabulary, so hosts interpret them.
4. Length units: CAL-03 uses millimeter, but the spec lists only mm; long forms of the other units, case sensitivity, plurals and spaces are unspecified (only millimeter/millimetre were added).
5. Number limits: only the integer refusal sentence is given (the >= form is from HOST_PARITY I06); float wording and the case of a GUI box that cannot hold the number are unspecified; the spec writes the key as limits.form_integer_max = N, the file uses limits.form_integer_max.value.
6. A Tools divergence: the notebook form's refusal of an integer beyond 2^53-1 does not use the 'L' must be <= 9007199254740991 template of 6.2 (NUM-02/03). Not fixed here.
7. Lookup limits (5.4) and the other SPEC 8 working decisions are default, not confirmed; N-D, downsample d != 1 and TimeIntervalOf/ZStepOf fixtures wait for Wave 2.

Notes

• The wheel-content test needs a recent setuptools (it skips with the reason otherwise; it passed here with setuptools 79).
• Hosts (Napari, Fiji, QuPath) are not changed; their loaders are documented only.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y2SN8SuEakTdiWy3F2WUqN