fix(ui): reliable session running indicator + swarm retry + version (#156, supersedes #159)

[warren618]

Lands the session running-status indicator scoped out of #159, completes the swarm-retry capability into something actually usable, and fixes the #156 version drift. Credit for the UI indicator and the swarm-retry origin goes to @ai7eam-dev (Co-authored-by on both commits).

1. Session running-status indicator (from #159, scoped)

The indicator flickered or vanished on reconnect / page reload / sidebar.

• useSSE: add attempt.created/started, message.received, session.created, tool_heartbeat, llm_usage to knownTypes so a mid-stream replay still surfaces the working state
• stores/agent: track streamingSessionId, preserved across switchSession so the sidebar spinner persists when navigating away from a running session
• Agent.tsx: handle attempt.created/started (re-affirm streaming on reconnect/reload), keep streaming alive on tool_heartbeat, persistent pulse bar, reword "Thinking…" → "Agent is working…"
• Layout: replace the dead-code status === "failed" dot with a spinning Loader2 for the streaming session
• index.css: pulse-slide keyframe

Dropped from #159 on purpose: the correlation fix (landed in #158), the SSE/LLM timeout change (landed in #157), and the agent/src/providers/llm.py default bump (protected module — needs separate discussion).

2. Swarm retry (HTTP + MCP)

#159 added a retry endpoint + an unused api.ts method but no usable trigger. Completed end-to-end:

• api_server: POST /swarm/runs/{id}/retry — reconciles a stale "running" run first, refuses a genuinely active run (409), else re-launches a new run with the original preset + user_vars
• mcp_server: retry_run(run_id) tool, so the list_runs → spot-failure → retry loop works from OpenClaw / Claude Desktop (the frontend has no swarm dashboard; MCP + HTTP are the real consumers)
• api.ts: retrySwarmRun client method (parity with cancelSwarmRun)

3. Version drift (#156)

The shipped 0.1.8 wheel reported 0.1.7 from a hardcoded constant. cli/_version.py now derives the version from package metadata, falling back to reading pyproject.toml directly — pyproject.toml is the single source of truth, no constant left to drift.

Tests

• test_swarm_retry.py (new): missing-run / running-refused / happy-path relaunch
• test_cli_version.py (new): pyproject fallback matches declared version, resolved-not-unknown, guard against reintroducing a hardcoded version literal
• test_mcp_regression.py: lock retry_run as a well-known tool
• test_security_auth_api.py: traversal rejection for the new endpoint
• Full suite: 2884 passed, 43 skipped; frontend tsc --noEmit clean + vite build passes

Notes

• No protected-area changes (src/agent/, src/session/, src/providers/ untouched)
• MCP tool count bumped 35 → 36 across the 5 READMEs. The ar/ko README prose still carries a stale 22-tool list — pre-existing localization debt, tracked separately, out of scope here.

Supersedes #159.