feat: native pytest plugin (#175)

[AKogut]

Second half of #175 — closes the issue.

Python could already report through JUnit XML, but that round-trip loses what matters most:

• parameter values get baked into the test name, so every variant looks like a different test (and never buckets under one base, undoing feat: L4 parameterized variants as buckets under one base test (#169) #189);
• a retry is indistinguishable from a fresh failure, so pass-on-rerun — the strongest signal the scoring model consumes — never arrives.

pytest-flakemetry reads pytest's own reports instead.

What it does

• Parameters travel as structured params (test_login[chrome] → title test_login, params: {browser: "chrome"}), so variants bucket under one base test.
• Retries are recognised via pytest-rerunfailures: fail-then-pass is reported as flaky.
• Suite from the class, file path relative to the rootdir, error message and stack from the report.
• Run context from the same environment variables the JavaScript reporters read (GITHUB_SHA, GITHUB_REF_NAME, GITHUB_RUN_ID, FLAKEMETRY_* overrides), so a Python run lands looking like any other — including the same idempotency-key scheme, so shards and reruns dedupe identically.
• Delivery is best-effort throughout — an unreachable endpoint or rejected token goes to stderr and is swallowed. An outage must never turn a green suite red. Stdlib urllib only; the sole runtime dependency is pytest itself.

Guarding against cross-language drift

The real risk in adding a second language is that the two implementations quietly diverge. So CI does more than run the Python tests: it runs a sample suite through the plugin and validates the emitted batch against the TypeScript ingestRunBatchSchema. The Python reporter cannot start emitting something the server would reject without the build going red.

I ran that check locally end to end before wiring it: a 5-execution batch with a failure and two parameterized variants validates clean.

Verification

• 11 Python tests, all passing locally on 3.9. pytester-based end-to-end coverage: pass/fail/skip mapping, parameterized variants carrying params, pass-on-rerun reported as flaky, class-based suites, no file written when nothing ran, and a dead endpoint leaving the run green (ret == 0). Plus unit tests for context resolution and idempotency keys (GitHub, local fallbacks, explicit overrides, shard distinctness, the ≥8-character contract minimum).
• New python CI lane on 3.9 and 3.12, path-filtered.
• pnpm turbo run build lint typecheck test — 61/61 green; pnpm format:check clean; docs build clean with dead-link checking on.

Notes

packages/pytest-flakemetry/package.json exists only so the contract validator can resolve @flakemetry/contracts as a workspace script; it is private and defines no build/test task, so the TS pipeline ignores it.