Fix DPC border handling with CFA-aware padding (closes #31)

[wuyiulin]

Closes #31. Port of the fix in 10x-Engineers/Infinite-ISP#50 (issue 10x-Engineers/Infinite-ISP#39).

Problem

apply_fast_dead_pixel_correction() relied on scipy's mode="mirror" to pad the raw
mosaic. Same-color neighbours are 2 px apart, so at rows/cols 1 and N-2 the mirrored
neighbour is the center pixel itself and img > max_value / img < min_value can
never trigger. The loop version apply_dead_pixel_correction() has the same bug via
padding() (np.pad(..., "reflect") on the mosaic).

Fix

• Add pad_cfa(): pads each of the 4 CFA sub-grids by 1 px with reflect
  (= 2 px on the mosaic), preserving Bayer phase.
• Fast path: pad with pad_cfa() before filtering, crop dpc_img and
  detection_mask back afterwards (so the debug count stays correct).
• padding() now uses pad_cfa(), so the loop version gets the same fix and both
  paths stay bit-exact. Happy to drop this part if you'd prefer to keep the scope to
  the fast path only.
• Detection/correction logic is unchanged.

Tests

New tests/test_dead_pixel_correction.py (75 cases):

• isolated hot/dead pixel at every border position (rows/cols 0, 1, 2, N-3, N-2, N-1)
• fast vs loop implementation bit-exact
• clean image unchanged, shape/dtype preserved
• pad_cfa() never mirrors a pixel onto itself and keeps Bayer phase

	this PR	main
pytest tests/test_dead_pixel_correction.py	75 passed	41 failed, 34 passed

On main, 40 of the failures are border defects left uncorrected (every position
with row or col 1 / N-2, hot and dead); the remaining one is the pad_cfa() test,
since the method doesn't exist there.

On in_frames/normal/ColorCheckerRaw_2592x1536_10bit_GRBG_100DPs_ISO100.raw
(dp_threshold: 80) the output is bit-identical before/after (95 detections),
and runtime is unchanged, so no regression on existing assets. pylint 10.00/10.

Note

Since the Reference Model is the basis for the RTL, the matching DPC block in
Infinite-ISP_RTL may need the same border handling to stay bit-accurate.

Pull request checklist:

• Issue created

ReadMe / Documentation: N/A (bug fix, no change to block diagram, results, or module description)

[marianadeem-10xe]

Thanks @wuyiulin. The fix and tests look good. Two changes are suggested in comments before merge.

[marianadeem-10xe]

One design suggestion: could you move pad_cfa() into util/utils.py as a plain function instead of using @staticmethod here? All helper functions in this repo are placed there, this design is common to all modules.

[wuyiulin]

Done. Moved pad_cfa() to util/utils.py as a plain function and updated the imports in the module and the tests.

[marianadeem-10xe]

padding function can be removed as it is unused.

[wuyiulin]

Removed. The loop-based apply_dead_pixel_correction() now calls pad_cfa() directly.

[wuyiulin]

Should I apply the same change (moving pad_cfa() to util/utils.py) to 10x-Engineers/Infinite-ISP#50 for consistency?

[marianadeem-10xe]

Should I apply the same change (moving pad_cfa() to util/utils.py) to 10x-Engineers/Infinite-ISP#50 for consistency?

Yes please. I missed it earlier. Apologies for the inconvenience.

[marianadeem-10xe]

Changes Reviewed.