chore: prune dead frontend components and stale doc refs

- delete unused components: StatusBadge, ThinkingBlock, ToolUseBlock
  (knip-confirmed, not imported anywhere outside themselves)
- remove unused Topbar re-export from shared/components barrel
  (Topbar is still imported directly by AppShell — component kept)
- drop unused @testing-library/jest-dom devDep (no imports anywhere)
- declare @shikijs/langs + @shikijs/themes as explicit dependencies
  (used directly by shiki.ts via deep imports — were only transitive)
- CLAUDE.md: drop ToolUseBlock/ThinkingBlock from the Conversation UI
  component chain now that the files are gone

make lint ✓, web tests 65/65 ✓, tsc clean. API-side tests need
Postgres up (unchanged, not touched by this cleanup).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: mariuspruvot

CLAUDE.md

@@ -49,7 +49,7 @@ infra/
 - **Container orchestration**: `container` module manages ephemeral Docker lifecycle, credential injection, result relay
 - **Skills as agents**: each skill is a self-contained folder with workflow definitions, mounted into containers
 - **SSE passthrough**: backend relays container output to frontend (no AI response generation in backend)
-- **Conversation UI**: frontend renders session output as a structured conversation with markdown (react-markdown + remark-gfm), syntax highlighting (shiki with JS regex engine), diff coloring, and collapsible tool_use/thinking blocks. Components: `ConversationOutput` (scroll container) -> `MessageBlock` (role dispatch) -> `MarkdownContent` / `CodeBlock` / `ToolUseBlock` / `ThinkingBlock`. Data flows as `StreamMessage[]` (structured content blocks) instead of flat text lines.
+- **Conversation UI**: frontend renders session output as a structured conversation with markdown (react-markdown + remark-gfm), syntax highlighting (shiki with JS regex engine), and diff coloring. Components: `ConversationOutput` (scroll container) -> `MessageBlock` (role dispatch) -> `MarkdownContent` / `CodeBlock`. Data flows as `StreamMessage[]` (structured content blocks) instead of flat text lines.
 - **Session persistence**: stream-json events are batch-persisted to `session_events` table (JSONB) during SSE streaming via `stream_and_persist()`. Completed sessions can be replayed from `GET /sessions/{id}/events`. The streaming pipeline is layered: `stream_events()` (raw tuples) -> `stream_and_persist()` (SSE + DB writes) or `stream_output()` (SSE only, for tests).
 - **Multi-turn via per-turn invocations**: `--input-format stream-json` exits after each turn (by design). The entrypoint runs `claude -p` for the first turn, then loops reading user messages from a FIFO and calling `claude -c -p` (--continue) for each subsequent turn. Each invocation emits its own `system` init + `result` events — frontend should expect multiple `system`/`result` events per session.
 - **Stream-json protocol**: containers emit NDJSON with 5 event types: `system` (init/retry), `assistant` (one event per content block — thinking/text/tool_use), `user` (tool_result), `result` (turn end + metadata), `rate_limit_event`. The `result.result` field duplicates the last assistant text — only display assistant events, use result for status only. No `--include-partial-messages` flag, so no `stream_event` deltas. A 6th type `error` (`{"type":"error","error":{"message":"..."}}`) is emitted by the entrypoint when setup fails (clone/checkout errors). The SSE `done` event includes both `message` and `status` fields — frontend uses `status` to distinguish `completed` vs `failed`.