ps-plate-03-verify: implement authoritative verification stage

[bryantharpe]

Implements the verification stage of the plate solver per openspec/specs/plate-solver/spec.md §Verification.

• Attitude: centroid-distance correspondence, coarse FOV from largest-edge ratio, SVD attitude solve, reflection rejection.
• Projection/match: diagonal-FOV catalog gather, derotate/project, in-frame trim, brightest 2·n_centroids cap, unique 1:1 matching within match_radius·width.
• False-alarm acceptance: binomial CDF test with Bonferroni correction; first accepted candidate wins.

Reuses math-core attitude/binomial/fov/pattern helpers and pattern-database nearby_stars query.

🤖 Generated with Claude Code

[bryantharpe]

review/judge — APPROVE on 65ebe2edf119

VERDICT: APPROVE
BLOCKING: NONE
NITS: verify.rs:103 — .expect("valid star index") on db.star_vector is a panic vector in a hot loop; consider propagating an error or filtering invalid indices upstream.
NITS: verify.rs:131 — comment says "brightness order is index order" but this depends on nearby_stars returning results sorted by brightness; if that contract is not guaranteed, the "brightest 2·n_centroids" cap may not select the actual brightest stars.
NITS: verify.rs — no unit tests for match_projected_to_centroids or largest_pixel_edge; both are pure functions that would be straightforward to test in isolation.
NITS: verify.rs:58 — reflection rejection is documented as part of this stage but the code delegates entirely to solve_attitude returning None; a brief comment noting that solve_attitude handles det(R) < 0 internally would make the delegation explicit.

---

Independent U3 review by the judge role (glm-5.2, Zhipu — different lineage
than every author seat). Verdict is on blocking findings, not a score. To
supersede with human sign-off (U4): judge-pr.sh override 80 "<reason>".

[bryantharpe]

review/judge — APPROVE on b025923d2477

VERDICT: APPROVE
BLOCKING: NONE
NITS: verify.rs:62 — db.star_vector(...).expect("valid star index") and verify.rs:80 — .expect("valid star index") use panicking expects on database lookups in the hot solve path; consider returning MatchResult::Rejected on None for robustness against malformed candidates.
NITS: verify.rs — no unit tests for the new module; coverage gate will assess, but dedicated tests for match_projected_to_centroids and largest_pixel_edge would strengthen confidence.
NITS: verify.rs:55 — reflection rejection relies entirely on solve_attitude returning None; an explicit det(R) < 0 guard would make the spec-mandated rejection self-evident in this module rather than implicit in math-core.

---

Independent U3 review by the judge role (glm-5.2, Zhipu — different lineage
than every author seat). Verdict is on blocking findings, not a score. To
supersede with human sign-off (U4): judge-pr.sh override 80 "<reason>".

[bryantharpe]

review/judge — APPROVE on b025923d2477

VERDICT: APPROVE
BLOCKING: NONE
NITS: verify.rs has no unit tests for the new verification pipeline (match_projected_to_centroids, largest_pixel_edge, ordered_pattern_pair are all testable in isolation)
NITS: verify.rs:33 doc claims "Reflections (det(R) < 0) are rejected" as a step in this module, but the rejection is implicit in solve_attitude returning None or producing only proper rotations — an explicit det(R) check or a doc clarification would prevent reader confusion
NITS: verify.rs:96 comment "brightness order is index order" assumes nearby_stars returns stars sorted by brightness; this invariant is load-bearing for the "brightest 2·n_centroids" trim but is not asserted or documented at the nearby_stars API boundary
NITS: verify.rs:53 candidate.edges[5] is assumed to be the largest sorted edge — the assumption is stated in a comment but not validated; a wrong edge ordering would silently produce an incorrect FOV estimate

---

Independent U3 review by the judge role (glm-5.2, Zhipu — different lineage
than every author seat). Verdict is on blocking findings, not a score. To
supersede with human sign-off (U4): judge-pr.sh override 80 "<reason>".