Roadmap

Known gaps against the v0.6.0 release.

Shipped in v0.6.0

Chessboard detector rewrite. calib-targets-chessboard is the invariant-first rewrite — precision-by-construction on a private regression dataset of blurred, lens-distorted frames (high detection rate, zero wrong labels). Types renamed: ChessboardDetector / ChessboardParams / ChessboardDetectionResult → Detector / DetectorParams / Detection.
Grid origin contract. Detection.target.corners is rebased to non-negative (i, j) with (0, 0) at the visual top-left (+i right, +j down in image pixels).
projective-grid standalone surface. The line / local-H validator, the circular-statistics helpers, and the BFS growth (behind a GrowValidator trait) live in projective-grid with no chessboard-specific dependencies. calib-targets-chessboard is the reference consumer.
Multi-component detection via Detector::detect_all / detect_chessboard_all. Same-board contract only; multi-board scenes are out of scope.
#[test] testdata regression harness. Per-image gates in testdata/chessboard_regression_baselines.json covering mid, large, small0..5, and puzzleboard_reference/example0..9.

FFI rewrite. calib-targets-ffi still mirrors the v1 chessboard param shape (with nested grid_graph_params / gap_fill / graph_cleanup / local_homography). Excluded from the workspace until the C-ABI surface is reshaped to the flat DetectorParams and the 3265-line src/lib.rs is split into purpose-scoped modules.
Seed hoist. The pattern-agnostic BFS grow already lives in projective_grid::square::grow behind a GrowValidator trait. The sibling find_seed + SeedCandidateFilter hoist is still in the chessboard crate — the seed finder’s 300-line chess coupling (Canonical/Swapped label split, axis-alignment classification at A, 2× spacing violation check) needs its own trait design pass.
example1 / example2 follow-ups. Two puzzleboard-reference images are tagged in the regression harness with ratchet_notes. example1 validation loop oscillates (needs either higher max_validation_iters or an accept-best-intermediate mechanism); example2 has a legitimate corner blacklisted by the edge-length cut under extreme view angle.

Degenerate axes (one axis with sigma = π) — current: drop the corner. Could a single-axis attachment pathway recover recall?
Seed retry policy — current: try the next-best seed. A blacklist-and-research scheme might catch genuinely-bad seeds earlier.
Distortion-curved lines — current: projective-line fit when ≥ 4 members, straight-line fallback. A true polynomial fit could absorb more distortion.
Multi-seed growth — current: single seed, multi-component via post-hoc booster. A first-class multi-seed grower could reduce the Stage-8 dependency.
Caller-provided cell-size hint — current: optional, mostly ignored. When could it tighten Stages 5–6 without compromising precision?

Contributions welcome.