Report content_rect from the CPU engine

[marcinz606]

The bug

Only the GPU engine writes content_rect. _on_render_finished does last_metrics.update(metrics) — a merge, not a replace — so the key kept the previous GPU render's value straight through a CPU render.

That is harmless while the two agree, and they usually do. With a crop active they do not: the GPU crops and then upscales the cropped region to the preview render size (gpu_engine.py:1673), reporting a rect for the upscaled buffer, while the CPU returns native cropped pixels. Measured on samples/20260619SP_EKTAR100_120_1_09_ME_4000PPI.tif:

gpu=True   buffer (1600,1184)   content_rect (0,0,1184,1600)
gpu=False  buffer (800,592)     content_rect (0,0,1184,1600)   <- inherited, 2x the buffer

main_window passes that rect to canvas.update_buffer whenever no border or paper aspect is set, which is the common case — the border branch is the only one that recomputes it. widget.py:362 get_pixel_rgb then maps normalized coords through the rect and clamps, so with the rect above every sample past 50% width pinned to the right edge: the white-balance picker and the colour readout sampled the wrong pixel. overlay.py:966,1425 map overlay coordinates through the same rect.

Reachable on any GPU-to-CPU switch with a crop active, including the automatic "GPU acceleration failed — using CPU" fallback, not only the manual toggle.

The fix

DarkroomEngine.process publishes content_rect too. No paper layout runs on that path — FinishProcessor/apply_carrier draw inside the frame rather than padding it — so the whole buffer is the picture and the honest report is None. Both engines now always write the key, so a render's metrics cannot inherit a stale one.

Verification

make all green.

tests/test_engine.py::test_the_engine_always_reports_content_rect seeds a stale rect into the context and asserts the engine clears it. Confirmed it fails without the one-line change and passes with it.

End to end on the real app, driving the exact switch that triggered it:

gpu buffer=(1600, 1184) content_rect=(0, 0, 1184, 1600)
PASS GPU still publishes its layout rect
cpu buffer=(800, 592) content_rect=None
PASS CPU render clears the stale GPU content_rect

Context

Found while chasing what looked like a CPU/GPU disagreement over crop_rect, flagged in #889. That part turned out to be a misreading on my side — both engines crop, they differ only in output resolution (correlation of the displayed picture +0.9995). This is the real defect that investigation turned up.