Phase B: routing domain core

[DocGerd]

Summary

Phase B of the implementation plan: the complete routing domain core — 13 tasks (B1–B13), each implemented and reviewed via subagent-driven development with per-task review loops, closed by a whole-branch review (final verdict: ready to merge).

• B1–B5 foundations: geodesy helpers; polar interpolation with no-go taper; wind grid space/time interpolation on u/v components; land/depth mask queries + 300 m navigable snapping; board & tack/gybe classification
• B6–B8 solver: isochrone routing with 45 s maneuver penalties and board-aware pruning; motor fallback (< 2.5 kn sailing → 6.5 kn motor legs, always flagged); collinear leg merging with mask+wind re-validation
• B9 two-rig orchestration (planRoute runs genoa + fock, recommends the faster; genoa on ties)
• B10 fast-check property invariants — land/shallow avoidance, time monotonicity, continuity, motor flagging, maneuver-count consistency, board-flip↔maneuver pricing — 25 seeded scenarios × 2 rigs, zero counterexamples
• B11 GPX 1.1 route export (rtept per leg + destination, ISO times, escaped desc with maneuver prefixes)
• B12 Web Worker protocol + typed RoutingClient: pure testable handler, transferred (not copied) mask buffer, per-rig progress throttling, promises settled on fatal/dispose (incl. dispose-race guard)
• B13 via-waypoint segmented routing — the router half of Movable via-waypoints on planned routes with automatic re-route #4 (Refs Movable via-waypoints on planned routes with automatic re-route #4; the draggable-marker UI half is E8, Phase E)

Test evidence

59/59 tests across 16 files (~209 s incl. property suite), typecheck + lint clean at a3a71a2. CI runs the same battery.

Notable review-driven hardening

Per-task reviews ran with adversarial fix loops; several defects originated in the plan's own code listings and were fixed additively: property-suite vacuous-pass guard (B10), CI-breaking unused import + empty-route guard (B11), promise settling on global fatal/dispose + dispose-race flag (B12), two borderline solver-test timeouts made explicit.

Deferred with named owners (tracked in the SDD ledger)

• Phase D: optional MaskMeta metadata fields (encoding/verticalDatum/sources) when mask.meta.json is first parsed; Plan is structured-clone-safe but NOT JSON-safe (documented on the type) — Garmin export (Garmin Boating (ActiveCaptain) route sync — import, and basic push of updates #3) needs a dedicated serializer
• E8: coincident-waypoint dedupe guard; multi-via E2E
• Phase E: clamp progress tMs monotone per rig at via joints
• F2: throttle-window E2E

Refs #4

🤖 Generated with Claude Code

[DocGerd]

CI note: the first run failed with 10 pure-timeout failures (no behavioral failures) — the Actions runner executes the isochrone solver ~6-10× slower than dev machines. The whole-branch review had triaged the timeout policy as "keep per-test explicit timeouts only", calibrated on local timings; CI evidence overturns that at per-file granularity. Landing: vi.setConfig({ testTimeout: 120_000 }) in the six solver-heavy test files (fast files keep the 5 s default so hang detection stays meaningful), property-suite cap 300 s → 900 s (observed 374 s on CI).