[Wenjie Fan (gggdttt)]

Current scope and immutable revisions

This supersedes the earlier no-knowledge comparison proposal. Preserve this PR as draft and unmerged. Both revisions below use the same knowledge-enabled orchestration and shared trace instrumentation; only the BCQuality ref differs.

Role	Engine commit	BCQuality commit	Content version
Shared baseline; required for the approved two-case diagnosis	73eb5c51fb7cdde453e3116778c4452c621ac9aa	41b0b7a39f890d904fd00f3a5c61582573c9adf8	1.9
Single-event-rule candidate; not the two-case diagnosis engine	0c9681369a8f9287d1d927282d8823c87fd0d001	d5aa1b8244ed04b281c45cce7c1ff598b2cffa2d	1.9

The candidate engine is a direct child of the baseline. Their entire repository diff is one ref: replacement in agents/ALReviewAgent/bcquality.config.yaml; all other Git blobs, including root prompt, trace schema, orchestration, tests and version metadata, are identical. Both retain Microsoft-only knowledge scope. Upstream microsoft/BCQuality#250 was verified open/unmerged at d5aa; its direct parent is exactly 41b.

Shared instrumentation base: 73eb5c51fb7cdde453e3116778c4452c621ac9aa. Its pre-instrumentation parent is frozen experiment 25817a39a67b9e883f1a8ea9624283ab2b41af9b; original main base is 9607dea1c84fcf660983a5e3791dece1fb2f1623. Earlier 5062/3c719/38180/ceaae/4b5c pins are historical only. No history was amended or force-pushed.

Bounded engine-owned trace

• Independent _review-trace.json v1; manifest declares trace_schema_version: 1. No BCQuality findings-report schema extension or skill/routing changes.
• Original leaf identity uses ordinal, original zero-based finding index and report hash. Bounded original candidate prose, supplied article refs, safe source location, severity, admission and deterministic normalizations support semantic diagnosis.
• Optional _root-dispositions.json permits only explicit model-declared kept/merged/downgraded/removed relationships and fixed reason codes. Missing, invalid or ambiguous relationships remain unknown. No guessed lineage, free-text reasons or exception text.
• final_index addresses _review-report.json#/findings, zero-based, before engine filtering. It is NOT a normalized finding index or a Bench F identifier. Trace does not establish downstream post-filter lineage; compare actual raw root and consumer artifacts separately.
• Limits: 1000 candidates/finals/dispositions, 128 diagnostics, 8 refs/indices, 1024-character candidate prose, 4MiB artifact; explicit truncation and body status. Defensive credential/control-character redaction is not a guarantee of complete secret removal. Retained prose requires the same access controls as final findings.
• The optional root prompt/output changes model input and may affect behavior. Both future comparison arms share them; this is NOT claimed literally behavior-invariant. Existing finding eligibility, severity thresholds, topology and defaults are unchanged.

Legacy engines without the manifest marker are handled as not_available_legacy by the downstream collector. Early failures before trace/manifest creation remain missing evidence. Historical frozen runs cannot be retroactively assigned leaf/root causes.

Validation and cross-repository evidence

• Actual engine functions generated four deterministic synthetic scenarios: explicit dispositions (including unknown), absent relations, 1002-to-1000 truncation, and fixed failure codes. No model calls.
• microsoft/BC-Bench#983 collector receipt confirms all four producer artifacts passed unchanged with schema validation. Collector revision: 546af054f12c0e0c82a8f71105b8a1c96ba3e992.
• Receipt helper hash caccf03569c8f6ea382bd115ee9fcde1461bfcc40f7e02539965a5228f8d1606 predates only a lint rename from $matches to $declarations. Reversing that rename reproduces the old hash exactly. Current helper hash: 54ff4851ca7e43effc5146339c9e0c89d337f3e29933fd31e95a63b95da84c94.
• Current helper regenerated all four scenarios offline into a separate directory; every output SHA exactly matches the collector's passed receipt, and all validate against the unchanged schema hash 3c6b6dd6abf0f0c32ec1201a1ecfb2996546c5151eccc6bbbbe91bbf0147c5de. Original evidence retained.
• Related orchestrator/metrics/trace suite: 345 passed, 3 failed. The three pre-existing schema-error-message assertions were independently reproduced on the pre-change baseline. Trace-only: 18 passed; separate entry-point/filter suites: 31 passed. Parse checks and diff checks passed; helper lint has no findings.
• Baseline exact-head CI: https://github.com/microsoft/BC-ALAgents/actions/runs/39920462701 (in progress at publication; not yet claimed green).

Existing shared CLI reproducibility fix

Optional caller-supplied COPILOT_AUTO_UPDATE is forwarded unchanged; no default is injected. Actual cleared-environment native probes with cached newer CLI retained showed unset => 1.0.102-1, explicit false => 1.0.88 on three probes. Actual startup accepted requested 1.0.88 and rejected deliberately mismatched 1.0.83. No model calls or global installation. This fix is common to both engine identities.

Execution hold

Only privacy013 and style012, once each, have diagnostic approval; they must use the 41b baseline above, including inspection of G03 versus G05 and G14-G16. Parent readiness review and explicit launch confirmation remain required. The full 145-entry knowledge comparison remains on hold. This PR preparation performs no model/evaluation dispatch, release, merge or Agent merge enablement.