feat(ai): Study Buddy endpoint — SSE + run history (AI-037b)

[mrviduus]

Phase 6 AI-037, slice b — the reader can now run the agent and watch it work. The panel UI is AI-038.

Changes

• POST /me/books/{editionId}/studybuddy — authenticated; runs StudyBuddyAgent on a highlighted passage and streams its progress over SSE: a step event per recorded step (index / kind / payload), a done event with the final answer (+ iterations + cost), or a terminal error event when the agent fails or exhausts its budget. The run is persisted (AI-036) on completion with the right status — a budget-exhausted run keeps its partial transcript. X-Accel-Buffering: no for the Cloudflare tunnel; client disconnect propagates untraced (no false trace); rate-limited (studybuddy, 8/min/IP — each run is several LLM calls).
• GET /me/studybuddy/runs/{runId} — returns a persisted run (scoped to the requesting user) with its step transcript parsed from jsonb, for the "show steps" view.

Design notes

• The SSE generator owns persistence: it resolves the scoped IAgentRunWriter from the request services (alive for the duration of the stream), accumulates the outcome, and writes one agent_run row in a best-effort terminal block — a failed write never breaks the response the client already received (surfaced as persisted:false on the error event).
• Tool DB access is safe under the streamed run: the dispatcher already creates a fresh DI scope per tool call (fix(ai): per-tool DI scope + streamed tool calls traced (AI-031b follow-up) #314), so parallel tools don't share the request DbContext.

Verification

• StreamRunAsync over the real agent + loop + scripted LLM: direct answer → step→done + a persisted completed run with the output; a never-terminating model → partial steps → terminal error + a persisted budget_exhausted run that keeps its transcript.
• Live-API integration (skip-friendly): no-auth → 401, authed empty passage → 400, no-auth GET → 401, unknown run → 404.
• Full TextStack.UnitTests (272) green; solution builds; dotnet format clean. The real streamed run (key + corpus) is exercised on prod / AI-039 eval.

🤖 Generated with Claude Code