E-6: render a slide as captured, and make its remaining controls honest

[seanharding]

Slides currently render nothing like the capture, and turning Normalize off doesn't fix it. This makes E-6 open faithfully, and cleans up several controls on that path that looked live but did nothing.

Measured against Photoshop, macOS Preview, Affinity and Darktable — which agree with each other, with NegPy the outlier.

The core change

With Normalize off, the Print stage is replaced by a transfer curve that is an exact identity at default settings, so the render is the capture:

• The camera's own colour matrix is applied. RAW decodes output_color=raw, so the buffer is in camera primaries. The print path never needed the matrix because it derives colour from measured density and a paper model; the transfer path can't, so the decoder's XYZ→camera matrix is inverted against the working space and row-normalised (dcraw's cam_xyz_coeff construction). A source without a matrix — scanner TIFF, JPEG — passes through.
• The exposure window is fixed, anchored to the decoder's white level rather than metered from the frame. This is what keeps a bracketed set rendering as a bracket: measured bounds make five exposures of one slide converge on the same render (spread 0.083), where a fixed window keeps them apart (0.544).
• A standard display rendering is applied — a fixed +0.7 EV baseline (darktable's shipped default for raw) and Narkowicz's closed-form fit to the ACES RRT + sRGB ODT. This is not optional: a bare linear-to-gamma encode is the one thing no consumer converter does, and measured against a Lightroom export it left midtones 1.37 EV dark with highlights 2.2× short of white. With it, the residual is +0.11 EV at the median.

Both terms are fixed constants, never metered — metering is what would make a bracket converge again.

Normalize now defaults to off

It was added as a rescue path — 0.9.10's changelog calls it "an optional Normalize step that tries to save expired slides" — but shipped on, so every slide met it as the general-purpose render.

On faded film it's the right tool. On a correctly exposed slide it isn't: the density runs to Dmax but only the top ~1.5 decades carry picture, so a window metered across the whole range (3.57 decades on the test frame) puts the median at 13% of the scale, which the print curve renders near white. Measured on a well-exposed frame: median luma 0.88 and mean saturation 0.14, against 0.70 and 0.29 for the same frame as captured.

The algorithm is untouched, so the rescue case is unaffected and it stays one click away. Only files with no saved edit change — e6_normalize is written into every record by to_dict, and the key shipped in the same release as E-6 mode, so no existing E-6 edit can be missing it.

Clamping the metered span was the alternative and was rejected: it fixes the look but costs the convergence Normalize exists to provide (bracket spread 0.083 → 0.136), and re-tuning one case degrades the other.

Controls that were lying

Three controls on this path moved a value that never reached the image:

• Crosstalk now declares its film process. A matrix describes one dye set, so profiles carry process and only appear — and only apply — in the film they were made for. A C-41 matrix can no longer be applied to a slide, or to B&W. The in-app editor gains a Process control and stamps new matrices with the mode in use, so a matrix you build for slides is one you can then select.
• Cast Removal is hidden outside C-41. The render already ignored it there — the solve needs the shadow and neutral-axis references, and both meters are gated to C-41 — so the slider was inert. It defeats the orange mask, a manufactured cast; a slide has no mask and its cast is the photograph.
• The Global/R/G/B selector is hidden on the transfer path, along with the White/Black Point sliders it scopes, which a fixed window ignores.

Conversely, Shadows/Highlights Density now work on slides. They existed, had sliders and persisted, but apply_transfer_curve implemented none of it — leaving four tone controls where a print gets ten, and no mid-sparing one at all. Their geometry is derived from the print's by tonal position, not by copying its density numbers: the curves don't share a scale, so the raw 1.50 would sit 64% of the way to black on a print but 50% here. The lift also tapers to nothing at the bottom of the window — a print gets that bound from paper black, this curve has no paper, and without it a lift walks the black point away (14% of a frame below 0.03 became 0.1%).

Bug fixed along the way

Crosstalk was silently skipped on the GPU for E-6 with Normalize off. The shader applied the unmix on the print branch only, so the Strength slider did nothing on the engine the app actually renders with — CPU moved 196/255, GPU moved 0. Every parity test ran with crosstalk gated off, where both engines agree trivially, which is how it stayed hidden.

Verification

• C-41 and B&W renders are unchanged, verified by SHA-256 fingerprints over float32 output buffers across a mode × paper × normalize sweep, on both CPU and GPU.
• GPU parity for the new shader is asserted on real hardware with engagement guards — each test checks the term is doing something on both engines, not merely that they agree, since a shader ignoring a uniform lane agrees trivially.
• 48 new tests in test_transparency_transfer.py and test_crosstalk_editor_process.py, plus additions to the sidebar suites. Full suite passes.

Behaviour changes for existing users

• New E-6 files open with Normalize off. Saved edits keep whatever they were saved with.
• B&W renders shift ~0.1%: it was receiving a Generic C41 unmix at the shipped default strength of 0.5, despite the sidebar having hidden the control there all along.
• Cast Removal and the channel selector disappear where they were already inert.

Related to #789 (another PR is coming to close that out later).