Add a toggle that renders the chat from qwen CLI's on-disk JSONL conversation
transcript (tailed live) instead of the orchestration projection. Read-only:
the transcript is parsed straight from disk, nothing is persisted on our side,
and approvals/turn control stay on the live ACP path.

Server (apps/server/src/ru-fork/qwen-transcript):
- paths/cliSettings: resolve the CLI runtime dir via branding (cliConfigDir,
  QWEN_RUNTIME_DIR, settings.advanced.runtimeOutputDir) — never a literal dir —
  to <base>/projects/<sanitizeCwd(cwd)>/chats/<sessionId>.jsonl.
- parse: qwen ChatRecord -> normalized TranscriptRecord union.
- QwenTranscriptService/Live: tail the JSONL (fs.watch + 1s poll fallback) and
  stream snapshot + append deltas over the orchestration.subscribeTranscript WS
  RPC. 75 tests, 100% coverage on the module files.

Web (apps/web/src/ru-fork/advanced-chat):
- merged tool-step flow: one card per tool use, correlating the assistant
  functionCall + tool_result + ui_telemetry, surfacing duration, diffStat,
  errors and cancellations; api_response telemetry folds into assistant latency.
  Telemetry is matched by event-name SUFFIX so a CLI rebrand keeps working.
- @pierre/diffs renderer for file diffs (raw Shiki fallback under
  ADVANCED_CHAT_RICH_DIFF); user bubble reuse + copy + image chip; theme-safe
  semantic tokens; heavy content collapses, nothing clamps.
- pending-op preview (ADVANCED_CHAT_PENDING_PREVIEW): un-applied edit -> jsdiff
  old->new diff; pending write_file -> collapsed content code block by file type.

Contracts: ru-fork/transcript.ts schema + WS RPC wiring.