refactor(metadata): finish wiring invoke formatter to pydantic union

[lstein]

Summary

• Completes the mid-flight refactor of InvokeAI metadata parsing by wiring invoke_formatter.py to the new GenerationMetadata2/3/5 discriminated union via a new InvokeMetadataView facade (the formatter was previously importing classes that no longer existed, so opening the drawer on any InvokeAI-generated image raised ImportError).
• Rebuilds the drawer HTML rendering with cleaner empty-field handling (suppress empty negative prompt, drop copy icon on empty positive prompt, per-column suppression in tuple tables with 1.0 fallback for gaps) and surfaces Flux Redux imageInfluence (e.g. "Medium") in the weight column when no numeric weight is present.
• Adds 33 regression tests covering view-level extraction and end-to-end HTML rendering for v2/v3/v5 (both ref_images and canvas_v2_metadata paths), plus Flux Redux, mixed weights, ref_images normalization, and empty-field edge cases.

What changed

Pydantic union plumbing

• New photomap/backend/metadata_modules/invoke/invoke_metadata_view.py — version-agnostic facade that exposes positive_prompt, negative_prompt, model_name, seed, loras, reference_images, control_layers, raster_images. All isinstance-dispatch lives here; the formatter stays ignorant of which v2/v3/v5 shape it's looking at.
• invoke_formatter.py rewritten as a thin HTML renderer over the view, preserving the existing drawer table layout. Falls back to <i>Unknown invoke metadata format.</i> on Pydantic ValidationError.
• _normalize_ref_images moved out of GenerationMetadataAdapter.parse and onto GenerationMetadata5 as a @model_validator(mode="before") (addresses a # MOVE THIS CODE TODO).
• metadata_modules/__init__.py re-exports (SlideSummary, format_exif_metadata, format_invoke_metadata) re-enabled — they had been commented out during the refactor, breaking metadata_formatting.py.
• Deleted invoke/invoke_metadata_abc.py — dead code, not imported anywhere under photomap/.

Drawer HTML tweaks

• Empty negative prompt → row suppressed entirely.
• Empty positive prompt → row present, copy icon suppressed.
• Tuple tables (loras / reference images / control layers): per-column suppression — drop columns that are empty in every row; drop fully-empty rows; missing weights in a surviving weight column default to 1.0 (InvokeAI's effective default).
• Flux Redux fallback: IPAdapter.image_influence / RefImageConfig.image_influence (values like "Medium") is displayed in the weight column when no numeric weight is available.

Modernization

• ruff check photomap/backend/metadata_modules/invoke/ --fix swept 227 typing.Optional/Union/List/Dict usages to PEP 604 / PEP 585 forms (X | None, list[T], dict[K, V]), plus one manual isinstance(x, (int, float)) → isinstance(x, int | float) fix.

Docs

• Added CLAUDE.md with orientation for future Claude Code sessions (architecture, commands, metadata subsystem layout).

Test plan

• pytest tests/backend — 82 passing (including 33 new tests in test_invoke_metadata.py)
• ruff check photomap/backend/metadata_modules/invoke/ — clean
• ruff check on all newly-authored / modified files — clean
• Manual verification in the browser: open the metadata drawer on a representative v2, v3, and v5 (both ref_images and canvas) InvokeAI-generated image and confirm the rendered table matches what the pre-refactor formatter produced
• Manual verification: open the drawer on a Flux Redux image and confirm the weight column shows the imageInfluence value (e.g. "Medium")

🤖 Generated with Claude Code