CI health: flake + speed analysis (last 40 runs)

Analysis of the last 40 ci.yml runs (2 more runs were a real tsc break mid-#128, excluded from flake math).

## Failure taxonomy

| Job | Genuine flake rate | Breakdown |
|---|---|---|
| iOS simulator tests | **7/35 (~20%)** | WaypointUITests announcement-while-backgrounded ×3 · RecordingUITests survives-backgrounding ×2 · KeyboardUITests ×3 (one = first-launch "Timed out attempting to launch app") |
| ci (web) | **4/38 (~11%)** | maplibre compass-realign ×3 (hardened in the companion PR) · record follow/track-up ×1 |
| iOS build | 0% | healthy |

The sim flakes are one family: **backgrounding scenarios under shared-runner load**. Reproduced the load-failure mode locally: 6 parallel cold WebGL boots fail before first paint; 2–3 parallel is stable.

## Time budget (sim job, median 14m)

XCUITests **366s** · build sim app 247s · boot+install 166s · everything else <1m. The test phase carries ~70s of hard `Thread.sleep` (RecordingUITests backgrounding windows) — the single biggest speed lever (30s windows likely shrink to ~10s without weakening what's proven; engine records at 1 Hz).

## Plan

1. ✅ Web: compass-realign hardened (companion PR). record:252 next if it recurs.
2. ⏭ Sim: needs macOS hands — **docs/ios-sim-runbook.md** (same PR) is the handoff: repro loop, ranked hypotheses, instrumentation, speed levers, ground rules. Est. impact: flake ~20%→<5%, job −3 to −5 min.
3. Boot/build overlap + cache-hit verification for the rust/derived-data caches (runbook §speed).

Generated from gh run history; job-level data in the session logs.