review-graph: adopt TypeSafe routing only after the shadow evaluation passes #74

Objective

Turn a successful #73 shadow evaluation into an optional production routing provider. A documented no-go result from #73 should close this proposal as not planned rather than force adoption.

Scope

• Use review-graph: evaluate TypeSafe/Jev routing in shadow mode before integration #73's measured applicability results and calibrated questions; start with coordinator-confirmed semantic additions and ambiguity handling.
• Preserve deterministic selection floors, shared ownership, independent review, proof identities, exact reuse, and all validation requirements.
• Keep normal coordinator routing available when disabled, uncertain, stale, or unavailable. Do not prune required reviews.
• Expose an explicit opt-in through the existing graph command surface. Keep the client dependency optional and support offline operation.
• Provide the process-scoped API launcher using local 1Password references and TYPESAFE_API_KEY. Do not publish keys or personal vault references.
• Bind model version, questions, evidence, decisions, latency, token usage, and configured cost limits to reproducible advisory records.
• Document the rollout and rollback path without changing the vendored upstream skill.

Deferred review-quality comparison from #72

If the proposed provider changes dispatched review work, evaluate that behavior before production adoption. Reuse #73's captured cases and measurement infrastructure; avoid building a separate full-graph live benchmark before the shadow experiment has justified adoption.

Run the normal and proposed workflows on identical captured source with the same model/settings and a declared four-slot agent budget. Execute all required specialist reviews and the conclusion-blind independent pass. Compare adjudicated findings and required coverage; investigate any loss of valid findings. Preserve exact publication/provenance records and observed source/context reads, coordinator operations, worker elapsed time, and orchestration time, with concurrency/overlap and unavailable measurements explicit. Exercise validation blockers and continuation through a truthful complete or partial/resumable result without discarding accepted audits or rewriting failed/blocked attempts.

#72's scripted protocol findings do not establish semantic recall. This live comparison is conditional on adoption changing dispatched work; it is not a new prerequisite for #73's unchanged-plan shadow experiment. If #73 records a no-go decision, close this proposal as not planned as already specified.

Acceptance criteria

• review-graph: evaluate TypeSafe/Jev routing in shadow mode before integration #73 records a positive adoption decision with reproducible held-out results.
• Disabled/offline/failure paths retain the current graph's coverage and readiness behavior.
• Enabled routing can add justified review work without silently removing any selected path or owner.
• If adoption changes dispatched work, retain the controlled full-review comparison described above, with required coverage, adjudicated findings, measured costs, and accurate validation/continuation evidence. Resolve any loss of valid findings before rollout.
• Credentials remain local and live calls are explicitly enabled.
• Repository checks and focused integration tests respect the no-Git-mutation policy, including temporary fixtures.
• Document observed benefits, limitations, and rollback instructions.

This issue represents a later implementation decision, not authorization to integrate Jev while #68, #69, or #72 remain unresolved.