[Mathih13]

PR checks usually take about 19 minutes. Uneven durable jobs, a five-minute Session expiry wait, superseded runs, and an unbounded dependency install delay contributor feedback.

This change cancels superseded PR runs, adds install and job timeouts, and distributes all Module test targets by recorded duration. The distribution includes time for each job's Gateway checks. New targets remain included with a default estimate. The Wasm build now uses the same features as the durable fixtures.

The Session test checks the five-minute production interval, then uses an Operator-only debug reducer to shorten its private schedule. It verifies two real scheduled cleanups and preserves renewed Sessions. The test takes 3.83 seconds locally with the Wasm already built. The production interval stays unchanged.

The auction regression also used an obsolete test path and passed with zero tests. Its corrected command checks that the exact test exists before it runs.

What you verified

• cargo fmt --all -- --check and workspace Clippy with all targets and features passed.
• Module library tests with debug_reducers: 1,569 passed.
• Gateway tests: 2,139 passed across all executed targets.
• The ignored Session expiry and exact auction regressions each ran one test and passed.
• python3 -m unittest discover -s scripts/tests: 6 passed, including complete, unique selection of all 82 Module targets.
• actionlint 1.7.12 passed.
• scripts/check-gateway-bindings.py: all 1,184 binding files match. SpacetimeDB schema extraction failed twice before the final check passed.

Local Rust checks used CARGO_PROFILE_DEV_DEBUG=0 and CARGO_PROFILE_TEST_DEBUG=0. Rebased onto current main; its intervening change only added a skill.

All nine GitHub checks passed on 95028a5a. The durable workflow completed in 16m18s. Its three jobs took 11m16s, 15m46s and 9m12s. This is one observed run, compared with the earlier PR median of 19m15s. CI ran the complete configured durable suite. The documentation push cancelled both jobs still running on the superseded revision, which confirms PR cancellation works.

Two independent adversarial reviews found no blocking defects. The second review confirmed the corrected testing instructions. An earlier attempt failed during private fixture startup with connection refused; the same failure occurred before this PR. The final run passed without a fixture code change.

What you did NOT verify

Long-term timing and cold-cache behavior remain unmeasured. No production deployment or migration was performed.

Model: gpt-6-astra, high reasoning effort. Harness: Codex in T3 Code.