feat(lif): extract objective lens metadata into audit + OME-XML

[ferrinm]

Summary

• New stdlib-only extractor zarrmony.metadata.objective.extract_objective(scene_xml) walks <ATLConfocalSettingDefinition> and <Objective> LIF elements and merges. Returns only the keys the LIF actually surfaced (nominal_magnification, numerical_aperture, immersion, model, working_distance_um); missing individual fields are omitted, and scenes with no objective info at all yield None (so the audit omits the objective key entirely rather than persisting an empty dict).
• Wired into _lif_scene_channels so both channel identity and objective ride one XML parse. Threaded through to:
  • attrs.zarrmony.per_scene[i].objective on the audit — a queryable structured record for downstream analysis.
  • A top-level <Instrument><Objective/></Instrument> in OME/METADATA.ome.xml plus per-image <InstrumentRef/> and <ObjectiveSettings/> — the standards-compliant surface consumed by napari, OMERO, and ome_types.from_xml.
  • Both per-scene and per-tile (lif_mosaic="per-tile") write paths participate. Plate + bf2raw are unchanged for this slice — those paths already skip LIF-specific channel extraction and are out of feat(lif): extract objective lens metadata (magnification, NA, immersion) into audit + OME-XML #52's scope.
• Objective extraction is deliberately decoupled from the channel-count check that declines omero identity on SizeC mismatch: a scene with a garbled channel list still has a valid objective and the audit shouldn't drop it. LIF DRY (Leica's air shorthand) maps to OME's Air enum; unknown immersion strings map to Other rather than being dropped.
• AUDIT_SCHEMA_VERSION bumped 6 → 7 (additive; consumers pinned to 6 should widen their pin).

Fail-closed hardening mirrors lif_channels / lif_tiles: billion-laughs, XXE, DOCTYPE, oversized/malformed input all yield None. Metadata never crashes a conversion.

Notes on the real LIF XML shape

The captured confocal fixture (tests/fixtures/lif_confocal_7ch.xml) surfaces the objective as attributes on <ATLConfocalSettingDefinition> — ObjectiveName, NumericalAperture, Immersion, ObjectivePos, ObjectiveNumber — with no separate <Objective> element and no WorkingDistance attribute. Two consequences worth flagging vs. the issue text:

• The bare Magnification="21" attribute on ATLConfocalSettingDefinition is total zoom (objective magnification × zoom factor), NOT the objective's nominal magnification. Recording it would silently write 21 for a 20x objective. The extractor prefers ObjectiveMag / NominalMagnification and falls back to parsing the NNx embedded in ObjectiveName — never Magnification.
• WorkingDistance isn't present in the fixture; the extractor still handles it (_fill_from_attrs), but its unit convention isn't cross-validated. The value is taken verbatim, and the audit key is working_distance_um — consumers with a real WD-bearing LIF should double-check their instrument's convention.

Test plan

• uv run --extra dev pytest -q → 355 passing (was 352 + 3 new wiring + 20 new extractor - 20 = net +23).
• uv run --extra dev ruff check src tests → clean.
• uv run --extra dev black --check src tests → clean.
• New unit tests in tests/test_lif_objective.py (20 tests): real fixture happy path, complete/partial/absent acceptance triple from feat(lif): extract objective lens metadata (magnification, NA, immersion) into audit + OME-XML #52, <Objective> sibling shape, ATL/<Objective> merge precedence, immersion mapping (DRY → Air, GLYC → Glycerol, unknown → Other, empty omitted), fail-closed hardening.
• New end-to-end wiring tests in tests/test_lif_objective_wiring.py (3 tests): each of the three acceptance cases drives convert(), asserts both the on-disk audit dict AND the parsed OME-XML instrument/objective/settings.
• Existing test_lif_per_tile.py schema pin updated (6 → 7).
• test_api_lif_wiring.py updated to destructure the new 3-tuple and to assert the count-mismatch case still surfaces the (decoupled) objective.

Closes #52

🤖 Generated with Claude Code