ci: expose test progress and nonfatal hang tracebacks

[zengleixin-amd]

Problem

PR #1476 has an unresolved CI hang. Current output does not provide enough live context to identify the blocked test or stack. This independent PR adds diagnostics only; it is not a root-cause fix.

Change

Only pyproject.toml changes: add -vv and -o faulthandler_timeout=180 to CI pytest arguments. Keep all existing tests, the 300-pass behavior, coverage options and thresholds, shards and workers unchanged. No timeout-killing plugin, workflow changes, or application changes.

Validation

• Read installed pytest 9.1.1 faulthandler and pytest-xdist 3.8.0 sources. The timer covers the per-test runtest protocol (setup/call/teardown) and does not terminate the process by default. It does not diagnose collection or session-level shutdown hangs.
• Isolated temporary two-worker probe, plugin autoload disabled and only xdist explicitly loaded: a 3-second call and a 3-second fixture teardown both pass without diagnostics and produce no stack dumps.
• With -vv -o faulthandler_timeout=1, both stack dumps arrive live around 3.25 seconds after process startup; both tests still pass around 6 seconds. Concurrent worker stacks can interleave.
• TOML parse and structural comparison confirm only those three argv tokens were added. git diff --check passes; focused Ruff lint/format checks pass. No project suite was run locally and no existing coverage files were touched.

Follow-up

Monitor this diagnostic CI run. If a long hang recurs, retrieve the available live job logs and stacks immediately rather than waiting hours. If it does not reproduce, report that explicitly; do not claim the root cause is fixed. No existing CI runs are cancelled and this PR does not change the #1476 branch.

🤖 Generated with Claude Code