fix(workflows): canonicalize FINAL_DECISION audit details to camelCase (PR1/3)

[kaicianflone]

Summary

PR1 of 3 from ~/.gstack/projects/consensus-tools-toolkit/ceo-plans/dashboard-zod-trust-boundary.md.

Aligns node-executor.ts:596-598 with the canonical camelCase emit shape already used by GuardEngine.persistAuditEvents (core/guard-engine.ts:111). Three audit-event producers now emit consistent FINAL_DECISION details:

Producer	Emit shape (after PR1)
core/guard-engine.ts:111	{auditId, decision, reason, riskScore, guardType} (already canonical, no change)
workflows/node-executor.ts:596	{runId, boardId, decision, reason, riskScore, guardType, consensusMeta} (was: risk_score, guard_type, consensus_meta)
sdk-node/chat-approval.ts:61	{runId, decision, approver, votesReceived, votesRequired, idempotencyKey?} (no risk fields by design — human approval, no change)

Why

Three rounds of /review on the previous fix/dashboard-error-handling branch (PR #34, now merged) kept finding the same trust-boundary class of bug. Round-3 finding D — "NaN% rendered when valid JSON has unexpected shape" — traces to the producer/consumer shape mismatch: dashboard reads risk_score (snake), some producers emit riskScore (camel), others emit risk_score (snake). Aligning producers first lets a single Tier-0 schema in PR2 validate canonically, then PR3 wires the dashboard through that schema.

Changes

• packages/workflows/src/node-executor.ts:596-598 — snake → camel for the FINAL_DECISION audit details (3-line change).
• packages/workflows/tests/node-executor.test.ts — new producer contract test that runs a guard node end-to-end and asserts the FINAL_DECISION audit details match the camelCase shape (with regression guards against legacy keys).
• packages/core/tests/guard-engine.test.ts — new producer contract test for GuardEngine.evaluate (was already camelCase; now locked in).
• packages/sdk-node/tests/chat-approval.test.ts — new file, 3 producer contract tests covering FINAL_DECISION on quorum, idempotencyKey propagation, and the VOTE_RECEIVED branch.

Note: guardResults runtime type and GuardResult return shape keep snake_case (risk_score, guard_type) per the plan-eng-review decision 7 — those are the runtime layer, distinct from the wire/audit shape this PR canonicalizes.

Test Coverage

• ✅ pnpm --filter @consensus-tools/core test — 72/72
• ✅ pnpm --filter @consensus-tools/workflows test — 35/35 (+1 producer contract)
• ✅ pnpm --filter @consensus-tools/sdk-node test — 33/33 (+3 producer contracts)
• ✅ pnpm --filter @consensus-tools/dashboard test — 36/36 (unchanged; dashboard tests don't cover the affected render paths)
• ✅ pnpm --filter ... typecheck — clean across all 3 affected packages
• ✅ pnpm dep-check — clean (225 modules, 423 deps)

Known intermediate state

After PR1 lands, RunDetailPage.tsx:107 and BoardDetailPage.tsx:189,196 will render NaN% for FINAL_DECISION events emitted by node-executor.ts (they read risk_score/guard_type snake_case). Pre-PR, those same render paths were already broken for FINAL_DECISION events emitted by guard-engine.ts (which is already camelCase). After PR1 they're uniformly broken; PR3 will land the consumer fix that closes both. PR2 (Tier-0 schema + parser) bridges them.

EventTimeline.tsx:256 already handles both shapes (payload.guard_type || payload.guardType) so it survives the intermediate state.

Test plan

• Build passes for core, workflows, sdk-node
• All affected packages' test suites pass
• Dashboard tests still pass
• dep-check clean
• PR2 follow-up: Tier-0 finalDecisionPayloadSchema in @consensus-tools/schemas + parseTypedPayload API
• PR3 follow-up: dashboard wiring + DriftBanner + component tests

🤖 Generated with Claude Code