v0.2.6 — rework detection v4: ranked, dust-proof

Detect-from-photo retrained on the hardest real-world case: a dusty board photographed straight off the machine.

Rework detection v4

v0.2.5's detector used absolute brightness thresholds — on a dusty, unevenly lit in-machine photo it flagged 40% of the board. Retrained against 16 fresh operator-marked ground-truth regions on exactly such a photo:

• A channel stretch is only suspect when it is dead for millimetres (contiguous near-zero run) and dark relative to its neighbourhood (15 mm local reference — absorbs lighting and dust gradients; floored so a fully dead zone can't declare itself normal).
• Suspect tiles are clustered and ranked by evidence — real defects cluster, photo noise scatters. Up to 12 boxes are proposed, most-evidence first, so the table order is the confidence order: review from the bottom, the last few are the doubtful ones.
• Validation on the dusty photo: 16/16 ground-truth regions found with 12 ranked proposals (v3: 13/16, buried in 90 flagged tiles).

Photo tips for best results: vacuum the dust off before shooting, get close (the detector wants ≥12 px/mm on the board), even light, no glare.

464 automated tests.