feat(net)!: extract stats publishing into moq-stats with compressed tracks

[kixelated]

What

Splits moq_net::stats into collection and publication, and adds compressed stats tracks:

• moq_net::stats is now collection-only. Producer is renamed Registry (Registry::new(exclude) / Registry::disabled()); the publishing task, Config, and frame types are gone from moq-net. The three duplicated counter shapes (RawCounts, TrafficTotals, BroadcastFrame) merge into one serde-both-ways Traffic type, and (SessionTotals, SessionFrame) into Presence, so the wire shape and the rollup are the same type. A new Registry::report() returns the per-(broadcast, tier) detail and prunes dead entries (the GC that used to live in the task); snapshot() is unchanged and never prunes.
• New rs/moq-stats crate (depends on moq-net + moq-json), with both halves:
  • Producer builds the registry itself (the config prefix is wired as the registry's exclude prefix, so the stats-of-stats cycle break can't be misconfigured) and drains it on an interval. Each track is now a pair: the plain .json track is written through moq_json::snapshot with delta_ratio: 0, compression: false, which is wire-identical to the old hand-rolled output (one full JSON object per single-frame group, same field order); the new .json.z sibling uses merge-patch deltas inside a group-shared DEFLATE window, which is a fraction of the bytes since successive frames are nearly identical.
  • Consumer yields typed TrafficFrame / SessionsFrame off either flavor (ConsumerConfig { compression }), plus parse_node_path(prefix, depth, path) and traffic_track / sessions_track naming helpers, so downstream aggregators (e.g. the moq.pro dash/edge slurp) stop hand-duplicating the wire types, track tables, and announce-path parsing.
• moq-relay: Cluster.stats is now the Registry; StatsConfig::build returns the moq_stats::Producer, which main.rs keeps alive for the run. No config surface changes.

Why

The stats frames are cumulative counter maps that barely change tick to tick: the best possible case for merge-patch deltas + a shared compression window, which moq-json already implements for catalog.json.z. moq-json depends on moq-net, so moq-net could not use it directly; splitting publication out of moq-net (which only collects now) fixes the dependency direction and deletes a hand-rolled subset of moq-json in the process (net -760 lines before the new consumer half).

Compatibility

• The plain .json tracks are byte-compatible with the old output; .json.z siblings are additive. Existing consumers keep working unchanged.
• Breaking pub changes are confined to moq_net::stats (rename + type merge) and moq-relay's Cluster::with_stats signature, hence the dev target.

Cross-package sync

• No IETF draft change: the stats broadcast is an application-layer JSON format published by the relay, not part of the moq-lite/moq-transport wire protocol, and is not specified in any draft (same as catalog.json.z).
• No js mirror: js/net has no stats module (its only "stats" are WebTransport getStats() calls); a @moq/stats consumer can follow up if needed.
• Docs: doc/bin/relay/config.md stats section documents the .z siblings; rs/CLAUDE.md crate map updated.

Testing

• The moq-net registry tests now run synchronously (no task to tick) and cover report()'s prune-after-final-values contract plus serde forward/backward compat for Traffic.
• The publishing tests moved to moq-stats largely as-is; new coverage: compressed siblings exist alongside plain tracks, and a producer→consumer round trip asserts the plain and .json.z flavors decode to identical frames across an update (exercising the delta path).
• just check and RUSTDOCFLAGS="-D warnings" cargo doc pass.

🤖 Generated with Claude Code

[sourcery-ai]

Sorry @kixelated, you have reached your weekly rate limit of 500000 diff characters.

Please try again later or upgrade to continue using Sourcery