MAINT Python CI performance

[Roman Lutz (romanlutz)]

Description

Several Garak unit tests repeatedly discover unrelated dataset providers, and the latent separator test persists far more cases than its assertions need. Linux-only JUnit uploads also leave Windows and macOS timing regressions harder to diagnose.

• Replace repeated provider discovery with function-scoped fixtures that load fresh, complete local corpora. Keep real per-test SQLite persistence, sampling, validation, and disabled-auto-fetch behavior. WebInjection tests seed their existing small corpus instead of fetching datasets whose values are then mocked out.
• Limit the latent separator test to 92 shared groups while checking every family/trigger pair across all 14 techniques. Keep the separate uncapped 4,266-case population fingerprints and full PromptInject matrix coverage unchanged.
• Use PR numbers for concurrency identity so same-named branches from different forks do not cancel each other. Upload uniquely named OS/Python/extras JUnit artifacts across the matrix, add a coverage-job report, honor JUNIT_XML, and print the 25 slowest test phases.

No production behavior changes. Preserve all 24 test combinations, three all-file pre-commit jobs, existing triggers and permissions, Linux-only result publishing, the four-worker loadfile scheduler, and the 78% overall / 90% diff coverage thresholds. Cache policy is unchanged: an incomplete local cache experiment did not justify changing it.

Tests and Documentation

Three interleaved before/after pairs used the same 117 test identities in five Garak modules, with uv 0.9.17, Python 3.12.13, all extras, and pytest -n 4 --dist=loadfile on Windows. Every valid run passed with no skips. The baseline and optimized runs used matching configuration, production imports, dataset contents, and parent memory fixtures.

Focused-suite wall time	Before	After
Pair A	193.73s	42.95s
Pair B	198.49s	29.10s
Pair C	181.47s	30.12s
Median	193.73s	30.12s

This is an 84.45% reduction for the local targeted suite, not a measured full-suite or hosted-CI reduction. The shared host had another workload; pairs were interleaved and counterbalanced. A preliminary snapshot run with collection errors was excluded entirely.

Ruff, formatting, type checks on the four changed Python files, actionlint, targeted workflow-contract checks, and commit hooks passed. The full suite and hosted matrix were not run. Updated the contributor test guide for JUnit paths and timing artifacts.

JupyText: N/A; no notebooks changed.