feat(ais_core): implement Secret-Redaction-Filter (#472, Issue #1d)

[SaJaToGu]

Closes #472 (Issue #1d — Wave 1b, Secret-Redaction-Filter)

Summary

Replaces the placeholder stub in ais_core/secret_filter.py with a full
pattern-based secret-detection implementation, plus comprehensive tests.

What's in this PR

ais_core/secret_filter.py

• 21 compiled regex patterns in SECRET_PATTERNS covering GitHub PATs
  (classic + fine-grained), OpenAI legacy + project, Anthropic, AWS
  Access Key ID + Secret assignment, Google API Key + OAuth, Slack
  tokens + webhooks, Stripe live secret + public, Discord bot tokens,
  JWT, Bearer tokens, PEM private key headers, env-var assignments of
  SECRET/TOKEN/KEY/PASSWORD/PASS/API_KEY/AUTH, plus Twilio SIDs,
  Mailgun keys, SendGrid keys.

• _COMBINED_PATTERN — single alternation regex built at import time
  from all SECRET_PATTERNS, used by redact_secrets for performance
  (one re.sub pass over the input instead of 21).

• redact_secrets(text) -> str — replaces all matches with [REDACTED].
  Pure function, no I/O, no logging side-effects.

• redact_dict(data) -> dict and redact_list(data) -> list —
  recursive deep-copy redaction. Strings get redact_secrets, nested
  dicts/lists/tuples recurse, non-string scalars (int, float, bool,
  None) pass through unchanged.

• Pattern-compile flags are passed as re.IGNORECASE | re.MULTILINE
  args (not inline (?imx)) so the pattern strings remain compatible
  with the combined alternation regex (Python's re rejects global
  flags in the middle of an expression).

Tests (42, was 11 smoke)

• TestPatternCatalog: >=20 patterns registered (21 actual),
  all instances of re.Pattern.
• TestRedactSecretsBasic: empty, no-secret passthrough,
  [REDACTED] marker, type-error on non-string.
• TestPositivePatterns: one positive test per registered pattern
  (21 patterns = 21 tests).
• TestRedactDict: flat, nested, list-of-strings, non-string
  passthrough, tuple-in-list, type-error.
• TestRedactList: strings, nested, type-error.
• TestFalsePositives: 15 safe samples (timestamps, branches,
  PR/Issue refs, normal log lines, file paths, URLs) + short-value
  rejection.
• TestPerformance: 1 MB redact under 500 ms CI budget. Test
  docstring documents the soft-target/realistic-local-runtime/CI-
  budget distinction (100 ms ideal / 200-300 ms local / 500 ms CI).
• TestIdempotence: redact(redact(x)) == redact(x).

Note on test samples

Slack-token + Slack-webhook positive tests use
secrets.token_hex(...) to construct fake token values at runtime.
The literal strings are never present in source, which both
documents the test intent ("a real-looking fake token") and avoids
GitHub push-protection false positives when the literal patterns
match GitHub's own secret-format scanner.

Performance

The combined alternation regex on Python stdlib re takes ~230 ms
for 1 MB of safe text locally. Per #472 review (commit 654+), we
agreed:

• No external regex package for 0.10.0 Issue #1d — Secret-Redaction-Filter pure helper #472.
• CI budget at 500 ms (intentionally generous to avoid flakiness
  on shared runners).
• Correctness + secret coverage > aggressive regex performance.
• Future optimization path (if ever needed): split into fast-path
  high-confidence prefixes (ghp_, sk-, AKIA, xoxb-, AIza)
  that run first, and slow-path (env-var, generic) only on demand.
  This would bring common-case latency well under 100 ms without an
  external dependency.

Tests

python -m unittest tests.test_ais_core.test_secret_filter -v

→ 42 tests, 0.277s, all OK.

LOC

File	Lines
ais_core/secret_filter.py	217
tests/test_ais_core/test_secret_filter.py	306

Out of scope (intentional)

• No wiring into ais_core.json_contract.success_envelope /
  error_envelope (planned for 0.10.0 GA, separate issue).
• No CLI integration.
• No production-log-corpus false-positive test (the issue
  referenced reports/runs/**/worker.log from the last 30 days,
  but those are gitignored per repo convention; we use a curated
  safe-samples list in TestFalsePositives instead).

Parent issue

Part of #468 (Release 0.10.0 — AIS Tooling Interface).