feat(assets): half-frame mode — two frames per scan

[marcinz606]

What

Half-frame cameras (Pentax 17, Olympus Pen…) expose two photos per 35mm frame; scans arrive as one file with two images separated by an unexposed gutter. This adds a Half Frame toggle (Files toolbar, mirrors RGB-scan end to end): each scan appears as two frames on the contact sheet, each with independent edits, bounds metering, thumbnails, sidecars, and exports.

How

• Identity: discovery expands each file into two assets with hash <base>#1 / #2 (negpy/services/assets/half_frame.py). Every hash-keyed store — file_settings, edit_history, file_marks, disk thumbnails, EXIF, render memo — becomes per-half with no schema changes.
• Independent bounds: the decoded buffer is sliced at the gutter before the pipeline, so full-frame metering is per-half metering. No shader changes; GPU caches keyed by the suffixed source_hash.
• Shared decode: decode/preview caches key on the unsuffixed hash (base_hash()), so both halves reuse one decode. Slices are copied — cached buffers stay read-only.
• Gutter auto-detect: 1-D column-luma heuristic with three gates — local-median deviation (immune to frame-brightness differences), extremal-vs-both-sides (rejects in-scene step edges), full-height uniformity (rejects textured features). Center-split fallback; the crop tool trims residue.
• Collision guards: per-half sidecars name.N.negpy; export stems name_1/name_2; add_files same-path guard matches per half; EXIF path-fallback rehoming excludes # hashes so it can't steal a half's edit.
• Toggling off re-collapses to whole scans; per-half edits stay in the DB and return on re-enable.

Validation

• 19 new tests (tests/test_half_frame.py): detector polarities/fallbacks/rejections, slice partition, identities, sidecar/export suffixes, discovery expansion, dedup guard, per-half bounds independence.
• Detector validated on all 7 samples/half_frame Pentax 17 scans: 6 split dead-center in the gutter; IMG428 (invisible gutter, night frames) correctly falls back to center.
• Headless app run (Xvfb): 7 files → 14 filmstrip entries, both halves of one scan rendered independently at half-width with distinct stats.
• make all green — 2022 passed.

Out of scope (v1)

Divider-drag UI, per-file split (mode is global like RGB-scan), RGB-triplet + half-frame combination (triplets stay whole).

closes #553

[github-actions]

Preview load metrics (github-linux-x64)

format	cold load	throughput
CR2	0.760 s	29.51 MPx/s
NEF	0.596 s	41.27 MPx/s
ARW	0.212 s	94.98 MPx/s
RAF	0.623 s	26.11 MPx/s
DNG	0.631 s	37.91 MPx/s

fixtures: rawsamples.ch · sha: 2f6e2b2 · full JSON

[ycjhjiangxiao-commits]

Can you extend this function to split film strip scan into multiple frames?