feat: SRT/WHIP resilience parity via a shared supervision layer (#20)

[joeblau]

Brings SRT/WHIP (SessionPublisher) up to the RTMP publisher's network-resilience, by extracting RTMP's supervision decisions into a shared, unit-tested StreamCore layer that both publishers use.

The gaps (SRT/WHIP had none of these)

NWPath supervision · queue-stall watchdog · per-path bitrate ceiling · frame-shedding admission (setQueueDepth was never called → admit() always true) · mic-stall failover · path-gated backoff (the old 1s→30s loop burned attempts while down).

Approach — hybrid (pure deciders shared, transport machinery per-publisher)

• StreamCore (pure, CI-tested, +27 tests): ReconnectBackoff, classifyPathTransition, WatchdogEvaluator, MicStallEvaluator, PathDebounce. Relocated VideoFrameAdmission + NetworkPathSnapshot here (public). The tests pin RTMP's exact constants/rules — including the frozen-queue-no-recycle (app-backgrounding) and metadata-flip-no-clock-bump guards.
• RTMPPublisher: delegates only isolated arithmetic to those deciders; control flow, socket-close ordering, and supervisors are byte-for-byte unchanged (−141 lines, no behavior change).
• SessionPublisher: gains the full supervision additively — NWPath supervision, a 2s watchdog, single-flight path-gated jittered reconnect (parks instead of burning attempts), per-path setPathProfile (re-seeded on reconnect), live frame-shedding, mic-stall failover, setMaxRetryCount(0). Recovery from a path debounce runs on a dedicated task so a route-return cancellation can't contaminate the parking reconnect loop.

Process

Built via an understand→design workflow and three adversarial multi-agent review passes, which caught and this fixes: a permanent-dead-air-after-dropout bug (confirmPathLost stranded a still-connected session) and a backoff-defeating task-cancellation bug (recovery ran inside a debounce task the route-return cancels).

Verification

• StreamCore: 96 tests pass (incl. new shared-supervision suites)
• Full app builds against the iOS 27 SDK
• ⚠️ Network resilience itself is device-only — it needs a physical iOS 27 device + real SRT/WHIP endpoints + induced drops (Wi-Fi↔5G handoff, dead-zone, mic-mute, congestion), which can't be exercised in CI. RTMP core is untouched + test-pinned; the SRT/WHIP changes are additive. Recommend on-device validation before relying on it.

fixes #20

🤖 Generated with Claude Code