feat(dewarp): auto-flatten only curved docs via curl gate v1.099

KakaoTalk multi-share goes through the batch path (ScanPipeline), which
had no dewarp, so curved/wrinkled forms stayed bowed. But applying UVDoc
unconditionally skewed flat docs (the v1.095 complaint). Resolved with a
self-correcting gate: keep UVDoc only when it actually reduces measured
curl.

- AutoDewarp.apply(cropped, force): measures curl (std of near-horizontal
  Hough line angles) on the perspective crop; if curl > 1.3 deg and UVDoc
  drops it by > 0.3 deg, return the UVDoc result, else keep the crop. So
  flat docs are never touched (no skew) and curved docs get flattened.
  force=true (editor "펴기" toggle) bypasses the gate.
- Wired into ScanPipeline (batch) and EditorFlatten (the "펴기" toggle is
  now a force override; default is the auto gate).
- Reviewer caught a blocker: HoughLinesP outputs CV_32SC4, so reading rows
  into a FloatArray yields all-zero (gate permanently off). Fixed to
  IntArray — without this the gate never fires.

Validated in the python spike: flat docs measure 0.00 deg -> keep crop;
curved forms 2.0-2.25 deg -> 0.00-0.77 deg -> UVDoc adopted.

Constraint: AutoDewarp never recycles its input; callers recycle the pre-dewarp bitmap only when the result differs; the gate keeps UVDoc strictly when it lowers curl.
Confidence: high (logic/memory) / medium (on-device curl thresholds need a real-device sanity check)
Scope-risk: module
Reversibility: trivial
Directive: on device, confirm a curved KakaoTalk form comes out flat while a flat receipt stays unskewed; if thresholds are off, tune CURL_MIN/CURL_GAIN.
Tested: assembleDebug + testDebugUnitTest exit=0 (22 tests); python gate validation on flat vs curved; reviewer pass after the Hough-type fix.
Not-tested: on-device curl measurement (Hough on real captures).
Related: v1.095 (dewarp made opt-in), user KakaoTalk batch report

🐙 Autopus <noreply@autopus.co>