[Product Gap] Ship a source-authoritative detector vertical slice with independent efficacy evidence

[seonghobae]

Buyer-visible problem

AppGuardrail can inventory and classify evidence supplied by callers, but the current issue-detection foundation does not yet prove that AppGuardrail itself acquires source-authoritative evidence and directly detects the underlying control failure. A buyer cannot distinguish a real detector from a typed wrapper around a Boolean, list, log regex, or signed upstream result.

Scope

Implement one bounded end-to-end detector vertical slice on current develop, using the historical #815/#813/#763 cases as candidate real-world replays. Select the highest-leverage case whose source system and expected outcome can be reproduced without privileged customer data.

The slice must bind:

atomic cause → control obligation → trusted probe/acquirer → independently acquired source identity → typed assessment → independent oracle → persisted evidence → control-plane/API result

Required contracts

• Add explicit probe_ref and acquirer_ref identifiers; no caller-provided pass/fail value may substitute for acquisition.
• Record repository/source artifact identity with SHA-256 and relevant revision metadata.
• Fail closed for unknown detector family, missing source identity, malformed source evidence, ambiguous multi-cause ordering, duplicate evidence, and stale evidence.
• Keep legacy envelopes readable only through a documented compatibility boundary; new production evidence must use the canonical source-bound schema.
• Integrate the detector through the same production path used by CLI/control-plane/API consumers.
• Preserve PII for authorized work rather than indiscriminate masking: use tenant isolation, least-privilege access, encryption, immutable audit, purpose binding, retention policy, and field-level authorization.

Test-first acceptance criteria

1. A RED test proves the pre-change product can accept or classify caller assertions without acquiring the authoritative source.
2. A real source fixture and an independently authored oracle prove true-positive, true-negative, malformed, unavailable, stale, duplicate, and adversarial cases.
3. Production-predicate mutation tests fail when the source acquisition, obligation binding, identity verification, or outcome mapping is inverted or bypassed.
4. The exact production entrypoint is exercised in a black-box test and a live exact-head audit.
5. Production statement and branch coverage remain 100%; all public modules, classes, functions, methods, and error types have complete docstrings.
6. No ignored/skipped efficacy or GPU tests are used to claim completion.
7. ADR, ARCHITECTURE, threat model, test strategy, operability/runbook, requirement traceability, and CHANGELOG are updated with APA 7th references to the governing standards.
8. Exact-head Tests, appguardrail-scan, SAST, CodeQL, dependency review, supply-chain checks, and independent review pass before merge.

Source preservation

PR #911 and branch agent/issue-detection-contract preserve the prior inventory/documentation prototype. Reuse only reviewed, current, minimal pieces; do not copy the 20k-line draft wholesale or treat its registry-generated fixtures as an independent oracle.

Exit condition

A buyer can inspect one detector and verify, from immutable evidence and reproducible tests, that AppGuardrail—not a caller assertion—observed the authoritative source condition and produced the decision.