[kixelated]

Ports main's session_fanout_threads case from moq-dev/moq#7493 onto release, so the pre-cut pairing (just bench origin/release, run from main) can compare #7493's contention fix. This is a bench-only backport under CONTRIBUTING.md's rule (moq-dev/moq#7581). It finishes what moq-dev/moq#7586 left for later. The quest PR is moq-dev/moq#7695.

Changes

On main the case lives in the split session/mod.rs, which doesn't build on release. So it's ported by hand into release's single-file rs/moq-net/benches/session.rs:

• fanout_threads: main's function, minus the Backend type parameter (release has only the model relay, and Model::PREFIX is ""). Same shape (1 publisher, 64 viewers, 400-byte one-frame groups) and the same case IDs: session_fanout_threads/{moq-lite-06,moq-lite-07-wip}/workers={1,16}.
• threaded(workers): main's multi-threaded runtime builder.
• Clock and MeasurementClock: main's round clock, reduced to Wall and Process. release has no SESSION_CPU thread-CPU mode, so the Thread variant is left out. Room gains a clock field, Wall by default, so every other case still times elapsed time exactly as before.
• Module docs: main's session_fanout_threads paragraph.
• rs/moq-net/Cargo.toml: tokio's rt-multi-thread dev feature, and libc as a Unix-only dev-dependency, as on main. Cargo.lock gains only moq-net's libc edge.

No shipped code changes. Public API: none. Wire: none.

Notes

• Merging this touches rs/**, so release-plz will propose a patch bump for moq-net, though only benchmark code changed. I bumped no versions.

Verification

All on release (338cb9b3e8) plus this change, Linux:

• cargo test --locked -p moq-net --bench session (Criterion's test mode) passes all 318 cases, including the 4 new ones.
• A short cargo bench run (--warm-up-time 1 --measurement-time 3) of session_fanout_threads measured 1.89 ms (moq-lite-06) and 1.80 ms (moq-lite-07-wip) of process CPU per round at workers=1, against 2.75 ms and 2.55 ms at workers=16. Those are the contention costs perf(moq-net): fanout subscribers stop contending the shared track lock #7493 removes on main.
• cargo clippy --locked -p moq-net --benches -- -D warnings is clean, with and without --features fuzz. So are cargo fmt --all -- --check, cargo sort --workspace --check, and cargo shear. cargo sort also warns about moq-cli and moq-tokio, but those warnings are already on release, and this PR doesn't touch either file.
• The case IDs match main's source (rs/moq-net/benches/session/mod.rs, fanout_threads), whose group name is {PREFIX}session_fanout_threads with Model::PREFIX = "", and whose rows are {version}/workers={workers}.

Review

Codex (gpt-6.1-sol, medium effort) ran a local quest review --base origin/release, round 1, on head 283a1bba417ed2173ae24d9b33a57478c86fd0b6. It reported no findings.

🤖 Generated with Claude Code

(Written by Claude Opus 5.5)

[kixelated]

Automated review by review (OpenAI)

Reviewed commit: 283a1bb

• P2 — Keep the multithreaded Tokio feature off wasm (rs/moq-net/Cargo.toml:39–62): The unconditional dev-dependency enables rt-multi-thread for wasm32-unknown-unknown, which Tokio rejects at compile time. This breaks the existing documented cargo test --test wasm -p moq-net --target wasm32-unknown-unknown path. Make that feature a non-wasm target-specific dev-dependency and gate the threaded benchmark as needed.

Backporting the benchmark is a reasonable way to compare releases; target-scoping the runtime feature keeps the change focused. Static review only; I did not run the benchmark or tests.

[pullfrog]

✅ No new issues found.

Reviewed changes — bench-only backport of main's session_fanout_threads case onto release, hand-ported into release's single-file bench layout; no shipped code, public API, or wire changes.

• Multi-threaded runtime: threaded(workers) builds a new_multi_thread runtime; rt-multi-thread added to moq-net's tokio dev features.
• Process-CPU clock: Clock { Wall, Process } and MeasurementClock { Wall, Cpu } ported without main's Thread variant or SESSION_CPU plumbing; Room gains a clock field defaulted to Wall, so every other case still times elapsed time.
• fanout_threads case: 1 publisher / 64 viewers / 400-byte one-frame groups, registered in criterion_group!, early-returning off Unix and under SESSION_ALLOCS/SESSION_EXPIRY_ONLY.
• libc: added as a cfg(unix) dev-dependency; Cargo.lock gains only moq-net's edge.

Verified locally: cargo check -p moq-net --bench session --locked is clean, and cargo test -p moq-net --bench session -- --test session_fanout_threads runs all four cases ({moq-lite-06,moq-lite-07-wip}/workers={1,16}), Success, exit 0. The case IDs match main's fanout_threads::<Model> (Model::PREFIX = ""), so the just bench origin/release pairing compares like for like.

  ^{｜ View workflow run ｜ Using deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏}