fix(filter): adaptive magic-pro levels keep faint documents readable v1.101

Comparing our batch results against CamScanner on the same forms, the
clearest everyday gap was not detection but the finish: Magic Pro's fixed
levels (black 35 / white 205) crushed faint carbon-copy / thermal forms
to near-white, while CamScanner kept the light ink readable.

- magicPro now derives the black point from the document's actual ink
  density (5th-percentile luminance + 10, clamped 35..120) and the white
  point from the background (80th-percentile - 5, at least black+60).
  Dense documents land near the old black=35 (unchanged), faint ones lift
  the black point to preserve light strokes instead of clipping them.
- Validated in the spike: on a real receipt/form the readable-ink ratio
  after the filter rose from 2.9% -> 7.7% and 6% -> 10.9% vs the old fixed
  levels, while normal-contrast docs look unchanged.

Constraint: levels span is floored at 40 to avoid divide-by-small; black point clamped to 35..120 so dense docs keep current behavior.
Confidence: medium-high (validated on real samples; user's exact faint originals not available)
Scope-risk: local (magicPro only)
Reversibility: trivial
Directive: if a specific faint form still washes out, send its original and tune the percentiles.
Tested: assembleDebug + testDebugUnitTest exit=0 (22 tests); spike readable-ink comparison.
Not-tested: on-device on the user's faint medical forms.
Related: ours-vs-CamScanner comparison; detection model training (SmartDoc) continues separately

🐙 Autopus <noreply@autopus.co>