[FrekiManagarm]

Summary

Closes the SFU-side instrumentation of LUM-553: per participant and per leg, jitter, loss, RTT, NACK/PLI/FIR counters and transport totals now reach telemetry.track_samples / telemetry.peer_samples as a per-second time series. This is Task 7 of docs/plans/2026-08-30-db-self-hosted-rust.md.

• str0m stats enabled. PeerConnection::new takes the sample interval (telemetry.sample_interval, 1 s by default). str0m disables stats by default, which is why jitter/loss/RTT were never measured.
• telemetry::sampler: pure delta arithmetic. There is no delta on the first reading, a counter going backwards clamps to zero, and jitter is converted from RTP clock units to ms (a zero clock rate gives None, never inf).
• Transport → session: new TrackStats, PeerStats and IceState events. Egress stats are ignored on purpose because they would double count.
• Session pump: records TrackSample, PeerSample, TrackCodec and ICE state/events.

Deviations from the plan

• Samples are keyed by occupancy, not connection (Task 6 review, finding 3), and are only written for published tracks. The new Telemetry::published_track never mints an id, so no orphan track_id is produced.
• The codec is recorded on the 1 Hz stats tick, not per packet. The only per-packet cost is one HashMap<Mid, _> lookup.
• The clock rate is derived from the media kind, as to_packet already does, so no per-packet registry is needed.
• str0m 0.23 has no ICE Failed state, so IceFailed has no emitter.

Overhead (cargo bench --bench sampling)

The instrumentation costs 20–24 ns per packet, which is 2.5–4 % of the 580–900 ns fanout and about 0.02 % of a core at 50 peers. Sampling itself costs ~60 ns per track per second and runs off the hot path.

Relation to the monorepo branch

origin/feature/lum-553 holds the same work done in the former monorepo (apps/sfu, unrelated history). It was never ported to this repo. This PR re-implements it on the current tree and ports its bench. That branch is left untouched.

Known gap

A peer switching rooms gets no TrackPublished in its new occupancy, so no track samples are written after the switch. This predates this PR and is tracked in LUM-708.

Test plan

• cargo test: 158 passed, including 8 sampler unit tests, a published_track test, and an end-to-end test in tests/room.rs where a real ICE/DTLS/SRTP session yields ≥ 2 non-zero video deltas with jitter in ms. That test fails when stats are disabled (checked).
• cargo clippy --all-targets -- -D warnings
• cargo fmt --check
• Run against Postgres (SFU_DATABASE_URL) and check rows in telemetry.track_samples