Hue Trim, and one home for the capture-side colour corrections

[marcinz606]

Three commits: a new Hue Trim control, a regrouping of the capture-side corrections it belongs with, and a rename pass.

---

1. Hue Trim

Why

A narrowband or odd-phosphor scanning light rotates hues rather than casting them, so yellows read orange and greens go olive. White balance cannot fix it, because a rotation leaves no grey to correct.

Measured on one negative scanned twice (a narrowband-ish panel vs a broadband white LED, same body, same frame, both cropped, default settings): |ΔH| held within 16–21° across CIELAB chroma 4 to 60+.

target chroma	median |ΔH|
4–8	16.7°
8–15	14.9°
25–40	18.1°
60+	20.6°

Flat in chroma is the discriminator: a colour cast would have shrunk with chroma (a fixed a*/b* offset barely turns a saturated colour), and a general channel mix would have grown. So the error is a rotation, and the correction is one.

What

process.hue_trim, degrees, ±30, 0 = off, rotating a*/b* about the neutral axis.

• Where: the scene-linear print, after the H&D curve and before the working OETF (the domain the reference measurement was taken in).
• Neutrals are fixed points, so it cannot fight the per-channel colour clip in Process. The two are orthogonal: that one places the grey axis, this one turns everything around it.
• Rides the exposure stage rather than owning one, so a slider drag does not re-run normalization behind it.
• Stays inside RenderIntent.FLAT: a light's hue error is a capture defect like the sensor unmix, not a look, so a digital-intermediate master should already be free of it.
• Sticky, like Linear RAW and Narrowband, because a light source is a property of the rig rather than the frame.

GPU parity

The mirror inlines rgb_to_lab/lab_to_rgb verbatim from lab.wgsl (WGSL has no includes) and rotates transmittance before oetf_encode, because the GPU encodes at the end of the exposure pass while the CPU stays linear to the end of the pipeline. Theta rides a hue vec4 appended to ExposureUniforms; at 288 B the block already spanned two 256 B-aligned slots, so the row costs no extra slot (288 → 304).

Scope

Slider only. The sky-anchored auto-detect is validated in principle (θ +18.6° on the affected scan, +0.61° on a known-good one, so it correctly does nothing when nothing is wrong) but deliberately left out: n=1, it needs an anchor in frame, and it wants a "no anchor, no correction" rule. Better judged once the manual slider is in use.

---

2. One home for the capture-side corrections

The sensor unmix, the crosstalk matrix and Hue Trim all correct the capture rather than the negative-to-positive conversion, and each has a different cause: camera CFA, film dyes, light spectrum. They now sit together in a Calibration section whose first block is Trichrome Calibration (the sensor unmix). Process keeps normalization alone. The persisted sensor section key is unchanged.

Crosstalk is not only about film dyes

The matrix operates on log density, and in that domain the film's dyes, the light's spectrum and the sensor's CFA all arrive as the same linear operator, so one 3×3 absorbs them jointly. That makes a profile a property of a whole scanning setup, not of a film stock. Tooltips and docs now say so and encourage tuning per rig.

Relatedly, the bundled matrices are now honest about what they are: datasheet-derived, not measured, and describing the dyes alone. They carry (approx) and state that they only tell the whole story for a true RGB scan (Coolscan-style, one band at a time) or a calibrated trichrome rig. Under a broadband light plus a Bayer sensor the capture adds mixing a dyes-only matrix cannot describe.

Profile provenance (type)

New optional type key, measured / tuned / specsheet-based, recording where a matrix's numbers came from, which is what decides how far to trust one. The Matrix dropdown groups under non-selectable headings, and the editor gains a Type control.

Two deliberate details: an unknown or absent type falls through to Other rather than vanishing from the dropdown, and an unrecognised value loads as tuned rather than the first combo entry, so a hand-written type is never silently relabelled as a datasheet claim. Anything the editor saves defaults to tuned.

---

3. Renames and cleanup

• The built-in profile is Generic C41, not Default, which said nothing about what the matrix is. Saved edits store the display name, so migrate_flat_config rewrites the old value. The render is unaffected (a None matrix still falls back to the built-in) but the dropdown would otherwise match no row.
• Hue Trim was missing from the copy-settings catalog; added, so it can be pasted between frames. Correctly absent from _BOUNDS_INPUT_FIELDS, since it acts after inversion and never feeds the meters.
• Corrected a stale CROSSTALK.md claim that the bundled gallery is copied into the user folder on first run. It is read live (ensure_user_dir explicitly does not seed), which is why these renames reach existing installs.

---

Tests

• tests/test_hue_trim.py: the op's invariants (neutrals still, chroma and L* preserved, opposite angles cancel, the requested angle recovered from real hue measurements) plus CPU/GPU parity at +18.6° and −12°.
• tests/test_crosstalk_profiles.py: grouping by type, and that grouping can never drop or invent a profile.
• tests/test_crosstalk_editor_type.py: provenance survives a save and a rename; unknown types fall back to tuned.
• tests/test_capture_unmix.py: the Default → Generic C41 migration, and that a user profile of another name is untouched.

One trap worth flagging: the parity fixture is deliberately chromatic. A grey ramp makes both the parity and the has-any-effect assertions pass vacuously, since a rotation barely moves low-chroma pixels. The first draft passed while the GPU was applying nothing.

Beyond pytest, the reorganisation was checked against the real MainWindow headlessly: every SLIDER_ATTRS path and tutorial target/section resolves, both sidebars sync, shortcut tooltips apply, and the grouped dropdown renders its headings with no profile lost.

make all green: 3329 passed, 2 skipped.

[thetalkingdrum]

Nice!