feat(ai): admin view-transcript UI for crew/agent runs (AI-045)

[mrviduus]

AI-045 — Admin "view transcript" UI (Phase 7)

Makes the multi-agent reasoning chain inspectable. Every crew run (crew.autopublish/crew.seo) and single-agent run (studybuddy) already persists an agent_run row with a nested-step transcript (researcher→drafter→critic→editor); this surfaces it in the admin app. No schema change, no new table — read-only over the existing Phase 6 agent_run.

Backend (2 read endpoints, mirror GetTraces/GetTrace)

• GET /admin/ai-quality/agent-runs?agent=&limit=&offset= — list, newest-first, agent filter = exact-or-prefix (crew. narrows to all crew runs), clamp 1–100. List projection omits the heavy StepsJson+Output and truncates Goal to 120 chars (guarded).
• GET /admin/ai-quality/agent-runs/{id} — detail with full raw StepsJson + Output, 404 on unknown.
• StepsJson passed through raw, parsed client-side (same pattern as TraceDetailDto) — no brittle second server-side schema.

Frontend (new Transcripts tab on AiQualityPage)

Modeled on TracesTab/TraceModal: filterable list (All / crew.autopublish / crew.seo / studybuddy) + pager → click row → modal. The modal parses stepsJson into the step tree — each sub_agent is a collapsible {stage} · {agentName} panel with per-step usage; the critic is special-cased + default-expanded, rendering its JSON verdict as score chips (factual_accuracy/tone/length/banned_phrases) + a severity-colored issue list (blocker red / major amber / minor gray). Defensive throughout — malformed/empty/non-JSON steps fall back to raw, never throw.

The casing contract is load-bearing (a mismatch = blank transcript)

DbAgentRunWriter serializes run.Steps with default STJ options (a direct JsonSerializer.Serialize, unaffected by the HTTP pipeline's web-defaults), so the stored shape is mixed: top-level AgentStep records → PascalCase (Index/Kind/Payload/At), the sub_agent payload anon-object → camelCase (stage/agentName/status/usage/steps), the nested AgentUsage record → PascalCase, inner llm_response payload new { text } → camelCase. The frontend reads each exactly. CrewTranscriptJsonContractTests (pure, no DB) serializes a real 4-stage crew.autopublish run through the same factory+options and asserts every key the UI depends on, with negative asserts so it fails loudly if anyone ever puts camelCase/web options on the writer. (QA-added; adversarial review confirmed the contract correct end-to-end, 0 P1/P2.)

Tests

• CrewTranscriptJsonContractTests (5, pure) — the write↔read casing contract.
• AdminAgentRunsEndpointTests (integration, live-server — same fixture as the rest of the suite, run by CI's docker job).
• dotnet test tests/TextStack.UnitTests → 407 pass; StudyBuddy set-equality green.
• dotnet format --verify-no-changes → clean.
• pnpm -C apps/admin exec tsc --noEmit + build → clean.

Notes

Per-description deep-link from the AutoPublish/SEO pages (by the runIds their crew-generate endpoints already return) is a deliberate fast-follow; this PR is the self-contained AI-quality tab. Admin-only (/admin/* auth) — Goal can carry user passages / book source, acceptable for the owner's audit surface. UI is build-verified; browser-check needs an admin session (owner-side).

🤖 Generated with Claude Code