Test strip: density x grade grid on the preview

[marcinz606]

What

A darkroom test strip for Print Density and ISO-R Grade. The Tone panel gets a Test Strip button that prints the frame as a 6×6 grid of real renders — density increasing left to right, grade softening top to bottom — so the diagonals read light-to-dark and soft-to-hard exactly like a split-filter test strip. Click the patch you like and its settings commit.

Choosing density and grade previously meant dragging two sliders and re-judging the whole frame after every move. This shows all thirty-six candidates at once.

The ladders are absolute rather than offsets around the current values, so a strip printed off one frame is comparable to the next, and they cover most of both sliders' travel:

• density 0.4 … 1.9 in steps of 0.3
• grade R55 … R180 in steps of 25

Both straddle the defaults and stay inside the ranges their controls accept — outside them the values clamp, and two rungs would render the same patch twice, so a test pins that.

No gridlines are drawn over the mosaic: it reads as one print, the way a real test strip does. Only the patch under the cursor is outlined, so it stays obvious what a click would take, and the rung matching the current settings is accented on its axis label.

How

Follows the zone overlay's split: pure NumPy in features/exposure/analysis.py, painting in overlay.py, session-only AppState flags. No WorkspaceConfig field, no migration, no pipeline stage, no WGSL. Grid size is derived from the two ladders throughout, so 4×4 → 6×6 was a two-constant change.

The one thing the zone grid didn't need is real pixels at other configs. Those come from a new RenderWorker.build_strip slot on the render thread's warm ImageProcessor: density and grade are absent from the analysis cache key, so the per-frame metering is reused and only the exposure stage onward re-dispatches. The strip always renders at preview resolution, never HQ — 36 full-res renders would take minutes, and each patch is shown at a sixth of the frame's width.

Three decisions worth a reviewer's eye:

• Picking leaves Auto Density / Auto Grade alone. The patches were rendered under them, so flipping them would render something other than the patch that was clicked.
• The assembled mosaic is soft-proofed once, not per tile. The transform is per-pixel, so one pass is identical and 36× cheaper. Without it, a proofed canvas would blow out — the canvas buffer is already sRGB-baked when a proof is active.
• load_file clears the strip, not only request_render: the navigate-back memo path repaints without going through request_render, which would leave a strip hanging over a different frame. Compare and flat peek clear it too — both want the canvas.

Verification

make lint, make type clean; 2638 tests pass. 31 new unit tests across ladder math, worker, controller lifecycle and overlay paint/click.

Headless E2E on samples/DSC00448.ARW, both engines (gpu_enabled True and False, no fallback):

• all 36 patches are distinct renders — no clamped duplicates at the ladder ends
• every patch is bit-identical (max abs diff 0.00e+00) to a full render at its own rung, so patch assignment and mosaic slicing are exact
• over the whole frame, more density prints darker in every grade row, and the softest rung is flatter than the hardest in every density column

mean luma per rung                    p90-p10 spread per rung
[[0.5769 0.4645 0.3568 0.2583 …]      [[0.9392 0.9269 0.8722 0.7417 …]
 [0.5550 0.4581 0.3647 0.2798 …]       [0.8364 0.8274 0.7645 0.6460 …]
 [0.5375 0.4542 0.3736 0.2995 …]       [0.7227 0.7003 0.6462 0.5714 …]
 …                             ]       …                             ]

Two measurement notes, both about the probe rather than the feature:

• Patch means are deliberately not compared to each other. Each patch is a different part of the picture, so its mean reports subject matter, not exposure; monotonicity is measured against full-frame renders over one fixed region.
• Contrast is strictly monotonic in grade for every density column except the densest. At density 1.9 a harder grade drives more of the frame off the paper's bottom end, so measured contrast peaks mid-ladder — real print behaviour, and the sort of thing a test strip exists to show. The E2E asserts the end-to-end trend for every column and strict monotonicity for the rest, and prints the exception rather than asserting it away.

[github-actions]

Preview load metrics (github-linux-x64)

format	cold load	throughput
CR2	0.771 s	29.11 MPx/s
NEF	0.595 s	41.30 MPx/s
ARW	0.215 s	94.02 MPx/s
RAF	0.643 s	25.30 MPx/s
DNG	0.644 s	37.15 MPx/s

fixtures: rawsamples.ch · sha: cc4aadf · full JSON