redteam suite: non-blocking follow-ups from #29 review

[stefanoferi]

Non-blocking follow-ups from the review of #29 (admina-redteam detection-efficacy suite). These do not block the merge of #29; tracking them here so they are not lost.

• Corpus size / reporting — corpora are small (37/26/11 positives) and per-lang/per-tag cells are 1–3 samples. Show numerator/denominator per cell and mark the output indicative; grow the corpora (especially non-EN) over time.
• Label provenance — labels are author-assigned in-repo. Add a one-line note that they are not third-party adjudicated.
• CLI not shipped — scripts/redteam.py is not in [project.scripts], so a pip-installed user cannot run the CLI the docs reference. Add an entry point (e.g. admina redteam) or adjust the docs.
• Gate fp invariant — the gate compares raw fp counts, sound only because the corpora are sealed. Add assert cur["fp_samples"] == base["fp_samples"], or compare rates.
• Corpus edge cases — loop-en-009 (counter-reset labelled loop), non-Luhn test cards, and no hard negatives in the PII FPR set.