Sharpen the pixels being exported, not the preview's

[marcinz606]

Root cause: sharpening landed on the wrong spatial frequency

apply_output_sharpening multiplied the USM radius by scale_factor = long_edge / 1600, so sharpen_radius = 1.0 became a σ ≈ 6 px blur on a 24 MP export. Measured through the real export path (61 MP frame, GPU tiled), as gain over the unsharpened buffer:

f (cyc/px)	0.05	0.125	0.25	0.375	0.475
before	1.78	2.14	1.58	1.59	1.59
after	1.05	1.27	1.83	2.10	2.19

The peak moved from 8 px structure to Nyquist. The export was never under-sharpened — it was sharpened at the wrong scale for its own pixels, which reads as contrast rather than sharpness at 1:1.

Radius is now in output pixels on both paths (lab/logic.py, gpu_engine.py, lab.wgsl), sharing gaussian_kernel_1d as before so CPU and GPU stay bit-matched.

The gate had to come with it

The L* noise gate was 1.5–2.0. At a true 1 px radius |L - blur| tops out near 1.0 (p50 0.20, p99 1.03 on a real scan), so the gate closed completely and the first cut of this change produced no sharpening at all. Rescaled to 0.25/0.33 — the old value over the ~6× it had been implicitly calibrated against. Picked against measured alternatives:

gate	Nyquist gain	grain gain
(0.50, 0.67)	1.81	1.18
(0.25, 0.33)	2.36	1.32
(0.15, 0.25)	2.47	1.34

Best detail-to-grain trade of the three.

Second fix: resized exports were not area-correct

The GPU tiled path downscaled with INTER_LINEAR, and the CPU path's INTER_LANCZOS4 is bit-identical to bilinear on an OpenCV downscale (measured maxdiff = 0.0 on 4.13 — no prefilter on a shrink). Both now pick INTER_AREA when shrinking. Affects PRINT / TARGET_PX only; the ORIGINAL default never resampled.

Ruled out

No hidden blur at defaults (only the lab USM touches pixels; CLAHE, chroma denoise, glow, halation, IR/dust, flat-field and carrier are all gated off). No textureSample anywhere on GPU. No proxy leaking into export — the source cache key includes fast_decode. Raw decode is clean: 16-bit, no FBDD, no median passes, no wavelet denoise, half_size never set on export. JPEG is explicitly 4:4:4.

Left open

Demosaic is AHD for Bayer, the softest option in the build (DCB, AAHD, VNG also available). A laplacian-variance sweep favoured AAHD ~2.6×, but that metric cannot separate detail from grain on a film scan and a flat-versus-detailed patch split did not resolve it. Wants a slanted-edge target before anything changes.

Also noted, not touched: _to_uint8_jit / _to_uint16_jit truncate instead of rounding, while the luma variants round. Half-LSB downward bias on RGB export, unrelated to sharpness.

Tests

make all green, 3613 passed. Two regression tests added — the USM must boost fine detail over coarse (test_lab_logic.py), and a layout downscale must match INTER_AREA (test_print_service.py). The relocation golden was regenerated; the characteristic-curve golden did not move, which is consistent since sharpen is not in the exposure kernel.

test_unmasked_half_is_bit_identical now pins sharpening off: a genuine ±3 px USM support carries the mask edge a few rows past the boundary on a 48 px test frame, which would hide the leak that test looks for.

Docs updated in the same change — PIPELINE.md (radius units, gate values) and USER_GUIDE.md (the preview no longer shows export acutance at fit-to-window; judge at 1:1).