PII detection regexes scale quadratically on digit-heavy input

Nobody has claimed this yet.

Observed while reviewing #3565 (present on main before that PR; timings identical at its merge-base).

`find_pii_matches()` in `agent-governance-python/agent-os/src/agent_os/credential_redactor.py` shows quadratic-ish scaling on pathological digit/IP-shaped input:

- `"1.1.1." * 8000` (~48KB) takes ~3.0s (phone/IP patterns)
- `"123-45-" * 8000` takes ~2.1s (SSN pattern)
- the `email_dots` secrets pattern takes ~5.4s on similar sizes

This path is reachable through the MCP gateway's PII detection on tool output, so a hostile tool can pad responses with digit soup to burn seconds of CPU per scan. Not a correctness issue (no bypass; the scan still completes), but worth bounding: input-length caps before the PII pass, atomic-group/possessive rewrites of the separator-heavy patterns, or a linear-time pre-filter (e.g. only run SSN/phone/CC patterns on windows around digit runs of plausible length).

A regression benchmark pinning worst-case scan time per KB would keep this from creeping.

Start in agent-governance-python/agent-os/src/agent_os/credential_redactor.py at find_pii_matches(), then reproduce the digit-heavy inputs and timings described here. Evaluate a bounded scan or regex/pre-filter change without reducing PII detection coverage. Add a regression benchmark that records worst-case scan time per KB and verify the scan no longer scales quadratically.

Written by the indexing model from the issue text.