feat(marshaling): activate dense signal in normal flow + densify mock (graph fidelity)

[ryan-johnson2]

Problem

The marshaling RSSI graph rendered at too-low fidelity — only the coarse path-1 stream (one scalar SignalChunk per node_data heartbeat → a handful of samples per node). The dense path-2 (current_marshal_data / get_pilotrace → SignalHistory, which signal_trace prefers) was built but dormant: RotorHazard only exposes a run's dense per-tick history for a saved heat, and the production staging loop never made a heat current (default practice-mode save_laps is a no-op).

Fix — two levers

1. Activate path-2 in the normal flow. The subtlety that drove the design: RotorHazard only records live laps for a node with a seated pilot once a heat is current — a selected empty heat rejects every crossing (server.py's pass gate pilot_id != PILOT_ID_NONE or current_heat is HEAT_ID_NONE). So selecting a heat at stage-time would silently break live recording (verified — it does). Instead:

• The heat runs in RH practice mode (live laps flow; the node interface still accumulates the dense history_values throughout).
• At heat-END the driver makes a savable heat current (add_heat → heat_data → set_current_heat, driver-thread-synchronous), then stops the race → DONE auto-triggers the transport's save_laps → race_list → get_pilotrace pull, which persists and returns the accumulated dense trace into the finishing heat's log.
• disarm now drives this graceful finish (stop + a dense-pull settle window routing the resulting SignalHistory into the right heat) rather than just clearing the armed slot.
• Heat selection is never emitted from the socket callback — heat_data is broadcast on every heat mutation, so an emit-per-event there floods and drops the link (a real drop I hit and fixed). The driver stashes the newest heat id (take_savable_heat) and selects once, on its own thread.

2. Densify the mock. The testkit CSV held the dense peak/nadir history columns flat, so RH's run-length dedup collapsed the history. node_csv/plan_csv now emit a textured per-tick peak/nadir envelope with distinct first/last times (two history entries per tick), while keeping the coarse node_peak (col 4) stable so the existing fidelity assertion still holds.

Adapter/transport: new RawHeatData parse + take_heat_ids/take_savable_heat + ensure_savable_heat. Also fixes a pre-existing non-exhaustive Event match in the rh_live test helper that blocked the live build.

Verification (throwaway dockerized RH — never :8080)

Through the production spawn_registry_bridge flow (rh_connect_live), running a real heat end-to-end:

coarse stream = 30 samples → dense history = 168 samples (full-fidelity upgrade
                                          activated by the normal heat loop)

The rh_signal dense gate (now driven via ensure_savable_heat, the production precondition): coarse 5 → dense ~80. The fidelity test still asserts exact coarse samples (== 180). The graph's SignalTraceView renders the dense trace unchanged (prefer-dense fold + canvas downsample).

Gates

• cargo xtask ci green (fmt + clippy + 78 test suites + binding drift).
• cargo xtask live: 16 targets pass — incl. rh_live, all 3 rh_signal, rh_connect_live (both tests, with the new path-2 production assertion), and every engine format e2e (the densified mock didn't regress any).
• Pre-existing, unrelated gridfpv-server full_event_live compile breakage on devel (router(state) vs EventRegistry) is untouched — out of scope.
• Did not deploy; did not touch :8080.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SL34wThSSbdutQoiLdJ2ZJ