[guijacquemet]

Stage 2 plan row PP-1 (MorphoBricks-Spec ADR 0002 AR8; contract §9.1 and §16), on main 43e3a93 (PP-C + PP-0). Part of v0.5.1. Do not merge before the independent review.

What changes

1. Fov.image_type, per position. "rgb" or "channels" (a unit with no image_type, experiment-v1 §4.4b), from the published description only, never from a header, never None on a Fov from fovs(). One source: FovIdentity.image_type (fov_identity.declared_image_type), which Fov and the resolver's FovSelection both carry. resolve_request now reads fov.image_type instead of looking the unit up a second time.
   • FovIdentity.image_type / FovSelection.image_type become str. They are "channels" where PP-0 gave None. RESOLVER_VERSION stays 1 because the resolver is not released before v0.5.1 and nothing consumes it yet. New constants: run_inputs_request.CHANNELS and IMAGE_TYPES.
2. Removed RunInputs.image_type() (breaking). It returned the manifest's run-level type from the command-line review, null on every described run, so it hid per-unit types. The manifest field stays as a record.
3. Partial-mask helper. make_described_run(..., masks=IDS, mask_images=PATHS) and write_described_tree(..., mask_images=) write masks only for the named images. testing.described_images(description) lists the tree's paths before anything is written. It is explicit only: no random subsets. These raise ValueError before any write:
   • a path outside the tree,
   • an empty subset,
   • a bare string,
   • a subset given without masks.
4. Memo (measured first). The description was already parsed once per content (description_consume._parsed). The cost was in read_checked, which decoded the whole mask_representation.json on every mask_check_status call. It now reads the bytes each call and reuses the parse for identical bytes (lru_cache). The result is identical, a rewritten file is parsed again, and RunInputs holds no memo and pickles as before.
   • Measured on a 10,000-position v09 run with Nuclei masks: 1,000 mask_check_status calls took 6.6 s before and 0.85 s after, with the same answers and the same resolve_request result.
   • Found and left alone: resolve_request still takes about 3 s there, and mask_check_status over all 10,000 positions about 12 s. Both come from linear scans per file (file_status, _mask_entry), not from the description.
5. Pinned API surface: tests/test_brick_api_surface.py covers the RunInputs public methods and their signatures, the fields and types of Fov, RunInputs, the resolver records and CleanRun, RESOLVER_VERSION, make_described_run, make_clean_run and described_images.

Tests

• New Spec module tests/test_spec_fov_image_type.py, added to spec_gate.EXPECTED_MODULES:
  • v28 is rgb on all 32 FOVs; v13 and v01 are channels.
  • v29's RGB unit and channels unit keep their own types through the identity. The runtime refuses that mixed layout (R_HETEROGENEOUS_LAYOUT_NOT_ENABLED), so the Fovs are checked on one run of each type.
  • v09 built by the helper with masks on 2 of 6 gives A_SCOPE_INCOMPLETE "present in 2 of 6" and rep1 partial through resolve_request, and gives a result equal to PP-0's hand-built tree.
  • The helper's refusals.
• tests/test_checked_status_memo.py: the parse is reused only for the same bytes; a mask step run after open is seen; the pickle is unchanged.
• PP-0's v09 test now uses the helper; its image_type is None assertions now expect "channels".

Gate

python tools/check.py: ruff check and format clean, code rules exit 0 (baseline not grown), 5069 passed, 4 skipped (filesystem and permission skips only). The Spec modules have 0 skips: tests/test_spec_*.py gives 1722 passed.

Downstream

CellPatcher main (347c61d) calls inputs.image_type() in run_open.py:365, normalization.py:188 and notebook_session.py:160. Its v0.5.1 pin (CP-4) moves to Fov.image_type. No other brick on main reads it: MorphoFeature, MorphoRelate, MorphoObjects, MorphoEmbed, fiberfeatures, MorphoCoverage and normalization were checked.

---

Generated by Claude Code