HOK-2358: Wrap workflow commands as native tools

[timogilvie]

Summary

• Wraps local Wavemill workflow commands as native runtime tools for review, routing, issue expansion, and stage-result recording.
• Adds structured success/error results, transcript metadata, and stage-artifact recording so workflow command failures preserve diagnostics.
• Makes stage-result writes idempotent updates and covers success, failure, and retry behavior with fixtures.
• Addresses self-review follow-ups for schema docs, UTF-8 truncation, expand_issue deps typing, route_task model-pool passthrough, and import ordering.

Changes

• shared/lib/native-agent/workflow-tools/command-tools.ts and tests: native wrappers for review_changes, route_task, expand_issue, and write_stage_result.
• shared/lib/native-agent/workflow-tools/linear-tools.ts and tests: Linear-facing tool execution and shared expand_issue execution surface.
• shared/lib/native-agent/workflow-tools/contracts.ts, contracts.json, and docs: native workflow tool result/request contracts and schema versioning.
• shared/lib/native-agent/workflow-tools/dedupe.ts, mutation policy/enforcer/record modules, and fixtures: idempotency, mutation recording, and fixture coverage.
• shared/lib/linear.ts and shared/lib/linear.test.ts: supporting Linear behavior used by the native wrappers.

Test plan

• npm test -- --run shared/lib/native-agent/workflow-tools/command-tools.test.ts shared/lib/native-agent/workflow-tools/linear-tools.test.ts shared/lib/native-agent/workflow-tools/contracts.test.ts (expanded to the repo test script; passed shell, unit, custom, smoke, and config suites).
• node --test shared/lib/native-agent/workflow-tools/command-tools.test.ts shared/lib/native-agent/workflow-tools/contracts.test.ts (68 passed).
• npm run typecheck (passed: 727 node tests, 47 workflow-router tests, 22 workflow-cost tests).
• Self-review tool run from the task worktree as required.

Self-review

• Iteration 1: exited 0, verdict ready; warnings fixed for docs schema version, byte-accurate UTF-8 truncation, and expand_issue dependency typing.
• Iteration 2: exited 0, verdict ready; warnings fixed for route_task modelsAvailable passthrough and dedupe import ordering.
• Iteration 3: exited 2 after Claude returned unparsable non-JSON on both attempts; required diagnostics were logged and PR creation proceeded per workflow instructions.
• No unresolved structured blockers were reported. wm:ready was not added because the final self-review run exited 2, not 0.

Routing

• Bootstrap route (.initial-route.json): planner claude-opus-4-7 depth medium, coder claude-sonnet-4-6 depth medium, reviewer gpt-5.5, review mode llm; provenance source bootstrap, input kind issue, routed at 2026-06-27T22:20:21Z.
• Actual planning execution (.planning-result.json): agent claude, model claude-opus-4-7, status completed, notes Plan approved by user, artifact plan.md.
• Recommended after expansion (.post-expansion-route.json): planner claude-opus-4-7, coder gpt-5.4, reviewer gpt-5.5, routing mode hokusai, provenance source expanded from the task packet.
• Active remaining route (.routing-complete / .phase-config.json): planning claude on claude-opus-4-7, coding codex on gpt-5.4, review codex on gpt-5.5 in llm mode.
• No routing.jsonl artifact was present, so runtime execution telemetry lines were not available.

Refs HOK-2358.