Product gaps: review depth, full-scope Strix, evidence-based fast routing

[seonghobae]

Product/control-plane gaps preserved from superseded traceability work

Protected main@9b57e4bb95b1a6efe9976a208fe7ca2c0d36dfec has already absorbed the #1507 multi-hour Noema/review-lifecycle changes, so the older #1535 document branch can no longer be merged as written. Preserve its still-valid buyer/control-plane goals here as focused implementation ownership rather than keeping a stale PR alive.

G-Review-Depth

OpenCode and Noema do not yet have a measured review-quality parity contract against strong independent reviewers such as CodeRabbit/Devin. Define a reproducible evaluation set and measure finding recall/precision, severity calibration, false-positive rate, exact-location validity, and repair usefulness on the same PR corpus. Audit opencode.jsonc, review prompts, scripts/ci/noema_review_gate.py, and contextual-orchestrator reasoning/test-time-compute settings. Do not optimize to finding count alone or invent a heuristic threshold; select acceptance criteria from measured data and research, then record the design and ablation evidence in ADR/doctoring and docs/product-technical-gap-baseline.md.

G-Strix-FullScope

Verify the actual trusted Strix scan scope on protected main. The product directive is that the required security review must be capable of whole-codebase analysis, not merely a changed-file slice that can permanently miss pre-existing exploitable paths. Design the full-scope path without weakening exact-head binding, artifact evidence, provider-unavailable fail-closed behavior, or source/credential isolation. Measure runtime and resource demand with real repositories; multi-hour scans are acceptable. Add current authoritative security/operations references and exact-head canaries.

G-FastModelRouting

Review/repair traffic through contextual-orchestrator should prefer models that are both capable and empirically fast for the task, rather than using provider/model-name lists or hand-written rules. Use structured runtime evidence keyed by the real deployment/model-group identity, including latency, terminal availability, capability, cost/privacy constraints, and later-success recovery. Route by measured evidence with ablation, not provider-name/model-id heuristics. The implementation owner may be ContextualWisdomLab/contextual-orchestrator; central .github should consume the versioned policy rather than duplicate it.

Acceptance

• current authoritative research/standards and provider docs recorded in APA 7 style;
• reproducible evaluation/telemetry schemas rather than rule-of-thumb constants;
• current-main implementation PRs in the repository that truly owns each boundary;
• complete statement/branch/docstring/edge-case coverage for owned code;
• exact-head required security/review evidence and no routine bypass;
• docs/product-technical-gap-baseline.md updated with current implementation and measured status.

Queue/capacity work remains separately owned by #1531.

[seonghobae]

Fresh owner-path evidence (2026-09-02): .github#1681 exact head e80fcc3daa72f93adb55116e155031c499dae9b0 makes categorical finding confidence = high|medium|low mandatory in the required Noema gate and derives it from adversarial_validation probe strength. That conflicts with the current Noema owner contract in noema#535@af55fad6fe0eddb1f90795509471bb96efb0576a, which explicitly rejects uncontracted Finding/ReviewVerdict fields, no longer publishes categorical confidence, and prohibits Noema from creating a second confidence policy outside contextual-orchestrator.

Treat this as G-Review-Depth RED, not a cosmetic schema choice: the same organization review path currently has two incompatible authorities for confidence semantics, and #1681 introduces an uncalibrated ordinal label into a required gate without the measured evaluation/ablation evidence this issue requires.

Acceptance before #1681 can be admitted:

1. RED: current contract test must fail if central .github requires a categorical confidence field that the live Noema owner contract rejects/does not publish.
2. Establish one owner for any confidence/calibration semantics; do not duplicate it in .github and Noema. If confidence is retained, it must be a versioned contract consumed from the owning boundary and backed by the reproducible evaluation set required here (precision/recall, calibration/error, location validity, repair usefulness), not by hand-mapped probe-strength labels.
3. GREEN: required Noema review succeeds with one compatible finding schema end-to-end; malformed/unknown fields still fail closed; severity admission remains independent; no model/bot confidence is promoted to human approval or security/compliance truth.
4. Re-run exact-head central tests/coverage/docstrings plus the Noema contract canary on unchanged heads. Queued/stale/predecessor evidence does not transfer.

No source/docs/ref/PR-state mutation was made to either active writer lane.