fix(detect): DocSeg as fallback-only + real-corner extraction v1.098

User's batch results came out slightly warped/keystoned on ordinary desk
forms after v1.097. Cause: the new DocSeg candidate could win on easy
shots where DocAligner is already precise, and its quad came from
minAreaRect (a rectangle forced onto a perspective trapezoid) -> distortion.

- CornerCandidates: DocSeg is now fallback-only — it joins the candidate
  list only when no DocAligner candidate exists (the hard cases it was
  meant for). On normal shots DocAligner handles, DocSeg never competes,
  so no regression / no warp.
- DocSegDetector: extract the actual 4 corners via approxPolyDP (preserves
  the perspective quad), falling back to minAreaRect only when the contour
  doesn't reduce to 4 points. The gate (positive ratio, solidity, area)
  is unchanged.

Constraint: DocSeg contributes only when DocAligner produced nothing; quad corners come from the contour polygon, not a forced rectangle.
Confidence: high
Scope-risk: module
Reversibility: trivial
Directive: if bowed edges persist on some pages, get the originals and confirm whether the "펴기" dewarp toggle was on — that produces non-linear (bowed) warps and is separate from corner detection.
Tested: assembleDebug + testDebugUnitTest exit=0 (22 tests).
Not-tested: on-device re-scan of the reported forms.
Related: 2256395 (introduced DocSeg candidate)

🐙 Autopus <noreply@autopus.co>