Solver: real harbor-to-harbor routes return 'unreachable' despite connected mask (blocks Flensburg→Marstal acceptance)

[DocGerd]

Found by E4's first real-data end-to-end browser smoke (real Open-Meteo wind, real mask.bin, real polars):

• Flensburg → Marstal: unreachable
• Flensburg → Glücksburg (~6 nm, deep well-connected fjord): unreachable
• Trivial open-water hop with no coastline nearby: succeeds

The mask itself is NOT disconnected: verify_mask.py's flood-fill gate (PR #8) proves both harbor snaps reach the open-water body at 3.0 m. So the failure is in how the isochrone solver expands against real coastline geometry — never exercised before, because all Phase B tests use synthetic masks (open water, blob islands, wide channels) and the property suite generates blob scenarios.

Working hypotheses (to be verified, not assumed)

1. Step length vs channel width: isochrone time-steps at sailing speed produce ~0.5–1 nm candidate segments; inside a ~200–400 m-wide fjord head every candidate heading's segment crosses land within one step → all candidates die at the first expansion (segmentNavigable correctly rejects) → blockedDeaths → unreachable.
2. Heading quantization: the candidate TWA set may be too coarse to thread a narrow channel even where a segment could fit.
3. MAX_FRONTIER truncation (documented-silent, see PR Phase B: routing domain core #5 notes) discarding the few candidates that do survive in constrained water.
4. Snap-to-cell-center placing the start deep inside the harbor pocket where hypothesis 1 bites hardest.

Impact

Blocks the core product promise and the spec's Flensburg→Marstal manual acceptance test. Phase E UI work continues (rendering verified with open-water plans), but this must be fixed before Phase F acceptance.

Investigation in flight: Node-side repro against real assets, death-counter/frontier instrumentation, root-cause report with fix options before any solver change.

🤖 Generated with Claude Code

[DocGerd]

Root cause found (full diagnosis in the repo's SDD ledger): (1) SOLVER — first-expansion isochrone steps (507–1165 m at dt=300–600 s) exceed the Flensburg harbor arm's straight sections (~200–420 m wide); all 39 first-expansion candidates failed segmentNavigable at t=0 → frontier 1→0 → 'unreachable'. Heading quantization, MAX_FRONTIER, and snap placement were each ruled out with instrumented evidence. (2) DATA, correcting this issue's premise for Marstal: its snap pocket only connects to open water at ≤ 2.3 m safety depth (consistent with its PR #8 depth exception at 2.0 m) — 'unreachable' at the 3.0 m default is CORRECT there; it routes fine at 2.3 m (8.0 h, 38 nm). Fix in flight on fix/solver-real-mask: blocked candidates retry the same heading at dt/2–dt/8 with honestly shortened clocks; no post-hoc passes, navigability still query-time-only. Flensburg→Glücksburg now routes under 5 wind directions; Flensburg→Marstal committed as the spec acceptance test at its documented 2.3 m caveat.