Phase B2: cross-camera homography ghost predictions (H-PROJ / Path A)

[mandarwagh9]

What & why

Phase B2 — the headline feature of the 10× roadmap. Implements cross-camera homography ghost prediction: when one camera sees a person and another doesn't, project the foot point between views to render a green H-PROJ ghost. Previously this was entirely absent (no cv2.findHomography in the pipeline).

⚠️ Stacked on the chain (#9 → #8 → … → #4). Auto-retargets as the stack merges.

Changes

• New app/infrastructure/homography.py — HomographyEstimator: accumulates foot-point correspondences per camera pair, estimates H via cv2.findHomography + RANSAC once ≥ min_pairs, re-estimates periodically, projects points, and caps the pair buffer (keeps most-recent).
• World-model integration:
  • store per-camera foot points on WorldObject (camera_foot_points)
  • _collect_correspondences feeds the estimator from objects co-visible on a tick
  • generate_predictions now tries Path A (homography → HOMOGRAPHY) first, falling back to Path C (world projection → WORLD) when no homography exists for the pair
  • wired from the existing HOMOGRAPHY_* Settings
• README — flipped homography / H-PROJ from 🔭 planned to ✅ implemented (status table, feature rows, AR-overlay table, the deep-dive section header, and the edge-case caveat).

Tests (TDD)

8 new in test_homography.py:

• estimator: recovers a known ground-truth H (sub-pixel), identity, directional independence, no-H-until-min-pairs, buffer cap, same-camera ignored
• integration: generate_predictions emits HOMOGRAPHY when a homography + foot point exist; _collect_correspondences feeds the estimator for a co-visible object

Verification

• ruff clean · mypy clean (17 files) · 89 tests pass (was 81).
• Coverage 55% → 58% (floor 50%).

🤖 Generated with Claude Code

[mandarwagh9]

Superseded — all commits from this phase landed on main via the merge of #13 (the top of the stack, which contains every phase commit). Closing to tidy up; the work is fully merged.