fix(useGameState): re-arm SSE silence check on every message — a stream that goes silent after the first event must fall back to polling again (run-178 root cause); unit: pin NZAA→NFTF full-pipeline marker lons on stored dataset coords

CHANGELOG.md

@@ -5,6 +5,8 @@
 - **Tests.** Unit +6 in `lib/mapData.test.ts` (new "antimeridian marker unwrap" block on the real NZAA/NFTF pair): eastward crossing unwraps a mid-leg −179.5 to +180.5 against the origin's +174.8, westward (reversed pair) unwraps below −180, and — the key invariant — every sample of `greatCircleArc(NZAA→NFTF)` fed back as a WRAPPED snapshot position comes out exactly on that arc's own sample, so marker and line agree frame to frame; grounded planes keep their raw coordinate, a flying plane with a missing endpoint stays wrapped (no line drawn either), and non-crossing legs are untouched. E2E +1 in `e2e/map-interactions.spec.ts`: the starter is relocated to Auckland via a fewest-legs chain over `/api/airports` (each hop ≤ 3400 km — within the ATR72's guaranteed range even in worst-case headwind; the tank is topped for free with the `set-fuel` hook before every dispatch so auto-refuel costs nothing), then a hook-created NZAA → NFTF package is loaded and flown at scale 1000 while the marker's longitude is sampled from the planes source every ~500 ms (real time) until the server reports landing — no two consecutive samples may differ by ≥ 180° (the teleport), some sample must sit on the east world copy (lon > 180, proving the crossing was actually traversed unwrapped rather than skipped), and the sampled span must exceed 4°. The plane chains back home afterwards so the other specs start from the canonical state.
 - **E2E fix (CI run 176).** The new antimeridian test failed deterministically in CI — not a marker bug, but a query trap: MapLibre's `querySourceFeatures` only returns features in LOADED tiles, and GeoJSON sources load per viewport tile; the default camera sits over Europe while NZAA/NFTF are in the Pacific, so the plane feature never appeared in queries and the "flying" poll timed out (the same pattern passes for EDDF→LEMD because that leg is inside the default view). The test now fits the camera over both endpoints first via `fitToAirports` (copied from `e2e/hover.spec.ts`, including its ISS-120 unwrap so a ±180-straddling pair bounds the Pacific, not the Atlantic) before polling and sampling. Pollution guard: relocation + crossing now run in a `try/finally` whose cleanup ALWAYS runs — it waits for any in-flight leg to land (scale 1000), then chains the plane back to EDDF from wherever it is, so a failed poll can no longer leave the starter FLYING and 409-cascade into the eleven later specs' fly-home helpers (the run-176 cascade). The dispatch POST's status is now checked explicitly with a clear message instead of a bare `expect(...).toBe(200)`. Unit +1 in `lib/mapData.test.ts` pins the exact feature properties the e2e `starterMarker` matches on (`properties.id === 'plane-starter'`, `properties.flying === true`).
 - **E2E fix (CI run 177).** The antimeridian test still failed — again not a marker bug: by dispatch time the test's page had been alive ~3 min through the relocation chain, and under dind load the browser's state feed can go stale (the SSE fallback in `lib/useGameState.ts` only re-arms polling on stream error or a one-shot 5 s silence check at connect), leaving the planes source frozen on an old FLYING frame from a relocation hop — which fooled the "wait for flying" poll and made every sampled longitude a stale pre-crossing value (continuous, all ≤ 180, never > 180). The test now reloads the page after relocation (server state persists; on the fresh page only the crossing leg can ever produce a flying frame), and the sampling loop gains a freshness gate: a sample is recorded only if its marker position is within ~700 km of where the SERVER snapshot says the plane is (~5 s of flight at ×1000 — haversine is ±360°-periodic in Δlon, so wrapped server lon vs unwrapped marker lon compare directly), so a frozen frame can no longer pass continuity vacuously; `starterMarker` now also returns latitude for that check. Pre-existing flake fixed in the same run: `e2e/queue.spec.ts`'s chain-refuel lower bound loosened from `1.4 × required2` to `1.3 × required2` — under dind load the post-departure snapshot can land ~0.1 × required2 into leg 2's burn (observed miss: 0.07 %); a wrong top-up level still lands far below the bound, and the exact level remains pinned by the upper bound plus the rep-discounted cash assertion.
+- **Root cause of the stale-feed freeze fixed in `lib/useGameState.ts` (CI run 178).** The reload + freshness gate above were insurance; run 178 proved the feed can freeze even on a FRESH page — and the reason is now fixed at the source, not worked around. The SSE silence check was one-shot: `setTimeout(startPolling, 5 s)` armed once at connect, cleared by the FIRST event and never re-armed. After that, only a stream ERROR could restart polling — but a half-open connection that simply stops delivering data (observed under dind load; `/api/state` on separate connections kept answering fine throughout) fires no `onerror`, so state froze until reload. The timer is now re-armed on EVERY message: any 5 s gap in the event flow — whenever it happens, not just before the first event — starts the polling fallback (immediate catch-up fetch + interval), which stops again when fresh events resume. Healthy streams pay one set/clearTimeout per event; stalled ones degrade to the pre-v0.57.0 poll loop instead of freezing. The e2e keeps both insurance layers: the reload still guarantees a clean SSE connect at dispatch, and the freshness gate still fails loudly if any future stall outlasts it (a frozen source can no longer pass continuity vacuously).
+- **Tests.** Unit +1 in `lib/mapData.test.ts` ("NZAA→NFTF crossing leg — full pipeline on stored dataset coords"): runs the exact production pipeline on the STORED CSV coordinates (`loadAirportsGeoJson` → `positionAt` slerp sampled at 51 steps → `planesGeoJson`) and pins what a live browser source must show — marker lons monotonically increasing from +174.8 to +184.9 (the unwrapped NFTF), every consecutive pair < 180° apart, max > 180 (eastward crossing onto the world copy). This is the ground truth the e2e's `lon > 180` assertion encodes: the dataset crosses EASTWARD (NZAA +174.786 → NFTF −175.149, Δ ≈ +10°), so a correct pipeline MUST produce lons above +180 — and any e2e failure of that assertion is now provably a stale browser source, not the math.
 
 ## v0.57.0 — 2026-09-22 19:19
 - **Push updated state to clients over SSE (ISS-129).** Clients no longer rely on pure polling for live updates: new `GET /api/events` streams full game-state snapshots as Server-Sent Events — the current snapshot immediately on connect, then one after every sim tick (~1 s), so the UI updates push-driven even while idle. New module `lib/sse.ts` is a tiny in-process fan-out registry (`subscribe(fn)` → unsubscribe, `broadcast(text)` with per-subscriber try/catch) — the three process-wide singletons (this subscriber set, the sql.js DB promise (`getGameDb`), and the tick lock (`withLock`)) live on `globalThis`: Next.js can load a module more than once per process (instrumentation.ts and route handlers get separate module instances), so plain module-level state would give each layer its own registry/DB/lock, and background-tick pushes would never reach route-handler subscribers. The broadcast hook lives in `tickNow()` in `lib/sim/engine.ts`: after each locked tick on the singleton path it serializes via the SAME `getStateSnapshot()` that `GET /api/state` returns (no duplicated serialization) and pushes to all subscribers; failures are logged, never fatal. Action mutations reach other clients within one background tick (~1 s); the acting client still gets its immediate post-action snapshot from the POST response as before. The SSE route (`app/api/events/route.ts`) sends `Content-Type: text/event-stream`, `Cache-Control: no-cache`, `Connection: keep-alive` and `X-Accel-Buffering: no` (prod sits behind Traefik), plus a `: ping` heartbeat comment every 15 s so idle proxies don't drop the connection; snapshot-first-then-subscribe ordering guarantees events are strictly ordered by sim clock, and stream cancel unsubscribes + clears the heartbeat; broadcast pushes are wrapped in `data: <json>\n\n` SSE frames at the route (matching the initial snapshot). Client (`lib/useGameState.ts`): one `EventSource('/api/events')` per page — onmessage parses the full snapshot into the shared state (same shape as today); polling stays as an automatic fallback: if no event arrives within 5 s of connect, or on a stream error, the old poll loop starts (immediate catch-up fetch + interval), and it stops again when fresh events flow. EventSource's native reconnect is relied on — no custom reconnect framework. `/api/state` behavior is unchanged (still ticks-then-snapshots; polling remains fully functional).