Harden Python perf interop: mplex teardown, yamux throughput, and perf_test.py reliability

[acul71]

Summary

Python↔Python perf interop in the unified-testing / test-plans harness is failing or flaky on real benchmark stacks. This issue tracks hardening perf_test.py, mplex (tls+mplex) teardown, and yamux window/throughput so perf runs complete reliably while preserving or improving measured performance.

Problem

• tcp + tls + mplex: expected connection-closed errors during post-benchmark cleanup cause non-zero exit after benchmarks finish.
• ws + noise + yamux: slow large transfers were cut off (~36% through upload 9) due to listener lifecycle / timeout interaction; yamux recv-window batching caused excessive round-trips on 1 GiB transfers.
• Harness ergonomics: no local Docker-free runner for fast iteration on perf stacks and PY_YAMUX_* tuning.

Scope (implementation in PR #1337)

1. perf_test.py (interop/perf/perf_test.py)

   • Listener: peer-polling teardown instead of sleep_forever(); connect deadline only until first peer; consecutive empty get_live_peers() polls before shutdown.
   • Dialer: swallow connection-closed errors only after _benchmarks_complete.
   • PERF_LOCAL_ADDR_FILE for local runner address handoff.
   • Configurable timeouts (TEST_TIMEOUT_SECS, DIAL_TIMEOUT_SECS, etc.).

2. mplex / tls+mplex

   • Extended shutdown grace for tls+mplex listener/dialer paths.
   • Phrase-based shutdown error classification (short-term; follow-up: typed exceptions per Replace fragile string matching on exception messages with typed exceptions across all transports #1218).

3. yamux throughput

   • Go-like GrowTo hysteresis, per-read WINDOW_UPDATE credit release, bounded read chunking.
   • PY_YAMUX_* env knobs for perf A/B (documented in scripts/perf/README.md).
   • Unit tests: tests/core/stream_muxer/yamux/test_yamux_growto_hysteresis.py.

4. Local tooling

   • scripts/perf/run_local_perf.py + README for matrix/quick runs without Redis/Docker.

Expected outcome

• python-v0.x x python-v0.x (tcp, tls, mplex) perf interop exits 0 after full benchmark.
• python-v0.x x python-v0.x (ws, noise, yamux) completes all 10 upload/download/latency iterations on 1 GiB workloads.
• Local ./scripts/perf/run_local_perf.py --matrix passes for regression during development.
• Throughput on yamux stacks is not regressed vs pre-fix baseline (target: fewer window-update round-trips).

Test plan

• CI: tox interop + core green on PR fix(perf): harden interop perf for mplex teardown and yamux throughput #1337
• Local: ./scripts/perf/run_local_perf.py --quick -t tcp -s tls -m mplex
• Local: ./scripts/perf/run_local_perf.py --quick -t ws -s noise -m yamux
• Optional: full --matrix (8 stacks) before merge

Related

• Supersedes perf interop test fails without venv-managed tooling and timeout hardening #1257 (venv/tooling subset of this work)
• PR: fix(perf): harden interop perf for mplex teardown and yamux throughput #1337