#392 feat(engine): PII/financial data-class exfil rule (closes #321)

[Anonymous]

Originally created by: fu351

Slice

• Repo: doberman-core
• Feature / Slice: Arc 1 — [#321] PII / financial data-class exfil detector (core-basic half)

What this PR does

Adds PiiDataClassRule (engine/rules/data_classes.py, wired into BUILTIN_RULE_TYPES): checksum-valid structured personal/financial data in an outbound payload — payment card number (known issuer IIN + Luhn), IBAN (mod-97), dashed US SSN (SSA validity constraints) — co-occurring with an external destination steps up to AUTH in every mode. New reason code pii_data_class_egress, documented in docs/REASON_CODES.md.

Precision is the design center, per the issue: presence alone never escalates (the rule reuses the secret rule's _has_external_destination gate and _scan_strings traversal), and every class is checksum/format-validated so random digit runs (timestamps, ids, hashes) fail. Boundary decision (ADR 0076): public-checksum structured formats are core-basic; free-text PII, OTP classification, and anything statistical is enterprise via the Detector seam. Deliberately excluded, recorded as ceilings: undashed SSNs, one-time auth codes, free-text names/addresses.

AUTH, never BLOCK — a human may legitimately send payment data; the point is a human confirms it. Raise-only end to end.

Tests added (run in CI)

• tests/unit/test_rule_data_classes.py (14): each class fires on the co-occurrence; spaced/dashed separators; URL-query fallback with no raw args; command-egress payloads; the co-occurrence gate (PAN to a local file → PASS); Luhn/mod-97/SSA failures → PASS; digits embedded in longer runs (epoch-ns) → PASS; the matched value never appears in the explanation or any build_record() field; the rule returns only PASS/AUTH.

Security checklist

• [x] Fails closed on error / uncertainty
• [x] No secret, full file, or unredacted prompt logged or committed
• [x] Any guardrail/learning change is raise-only (no silent loosening)
• [x] Every BLOCK/AUTH carries reason codes + a human explanation

Edge cases covered / Deviations from plan / Risks introduced

• Local battery: full suite green across two runs (autodeny run for the bulk + the 11 approval-flow files re-run without the switch, both logged in test-logs/). Benchmark gate + detection corpus unchanged-or-better (raise-only addition).
• CHANGELOG also records the just-merged [#368] OTel sink in the Unreleased block.

[Anonymous]

Ticket changed by: fu351

• status: open --> closed