test: add coverage for agents API, agent service, task-run repo, meeting analytics, and SSE health payload

[JLRansom]

Summary

• Exports buildSystemPrompt and buildUserPrompt from agentService.ts as testable pure helpers (no behaviour change)
• Adds makeTestAgent and makeTestMeeting factory helpers to __tests__/helpers/test-utils.ts
• 5 new test files covering all gaps identified in the spec:

File	Tests	What's covered
__tests__/integration/agents-api.test.ts	8	GET/POST /api/agents — happy paths, validation (missing fields, invalid type enum), 404 on non-existent project
__tests__/unit/task-run-repo.test.ts	14	dbGetTaskRunsByAgent (filtering, isolation, ordering, shape) + dbGetRecentTaskRunsByAgent (time boundary, empty window, boundary-inclusive, cross-agent isolation)
__tests__/unit/meeting-analytics.test.ts	7	dbGetMeetingAnalytics — empty project, token/cost aggregation, daily grouping, concluded-only filter, project isolation, date range, null-field handling
__tests__/unit/agent-service.test.ts	13	buildSystemPrompt (role-specific content, distinctness), buildUserPrompt (context injection, directory, role-specific sections), double-scheduling guard in assignAgentToTask
__tests__/integration/stream-health-payload.test.ts	5	getStreamPayload().agentHealth — key presence, per-agent entry, AgentHealthMetrics shape, hasEnoughData: false for sparse agents, hasEnoughData: true with populated metrics for rich agents

Result: 161 → 213 tests, all passing. Zero new type errors.

Design decisions

• buildSystemPrompt/buildUserPrompt are exported with export function — smallest possible surface-area change to make pure logic directly testable.
• makeTestMeeting messages use meeting.id (post-spread) for meetingId FK, not the pre-spread local variable — avoids FK constraint failures when callers pass an id override.
• seedMeeting in the meeting analytics test calls dbUpdateMeetingMessage after dbAddMeetingMessage for token/cost fields — mirroring how production code sets stats after streaming completes.
• Stream health test anchors timestamps to Date.now() (not a fixed past date) so runs always fall within the 7-day rolling window regardless of when the suite runs.

Test plan

• npm run test:run — 213/213 passing
• npm run type-check — no new errors (5 pre-existing errors unchanged)

🤖 Generated with Claude Code