fix(supervisor): restore session spawn via in-process claude-runner + bridge

[finedesignz]

Root cause

Phase 09 (PR #86) gutted `ProcessManager.spawn()` and finalized every `session.start` with reason `legacy_agent_spawn_disabled`, so clicking Launch on a repo in Settings → Supervisor / Sidebar did nothing. The retired external CLI agent had the autonomous-loop bug, so PR #86 removed it — but no in-process replacement was wired up.

Reproduced live in `supervisor.log` at 2026-05-27 23:51:54Z:

```
[security] legacy_agent_spawn_disabled: ... session.start cannot proceed until the
in-process claude-runner lands. Run will be finalized as stopped.
```

Fix

Re-wire `spawn()` to construct a per-run `SessionBridge`. Each bridge:

1. Opens its own `/ws/agent` connection with `project_dir = repoPath` and `role = 'agent'` — mirroring what the retired external agent did from its own process.
2. Spawns `claude --input-format stream-json --output-format stream-json --verbose` directly via a new in-process `ClaudeRunner` (NOT the retired external npm CLI; canary still passes).
3. Translates Claude's stream-json events to the hub's agent-protocol shapes (thinking / text_delta / tool_use / tool_result / status / assistant_message / permission_request / user_question / agent_log).
4. Routes hub-originated `user_message` / `permission_response` / `question_response` / `cancel` / `shutdown` back into the runner.

Preserved from PR #86:

• All five security gates (sandbox-escape, not-git-repo, concurrency, duplicate-run, dangerous-skip cap) + audit-log row per decision.
• `CIRCUIT_THRESHOLD = 5 crashes / 10 min` circuit-breaker.
• `MAX_RESTART_COUNT = 10` restart cap — so a broken spawn can never loop forever again.

Tests

• Phase 09 canary `supervisor/test/no-legacy-agent-spawn.test.ts` — still passes (new code uses neither the retired package name nor the forbidden `--append-system-prompt` flag).
• `supervisor/test/process-manager.test.ts` rewritten to use the new `bridgeFactory` hook in place of the deleted `spawnImpl` (Bun.spawn) hook. Asserts: session.start constructs the bridge with correct opts, gates still reject before bridge construction, dangerous-skip is correctly threaded through, audit log captures allow + reject decisions, crash path schedules restart.
• Full supervisor suite green: `49 pass / 0 fail / 7 files`.

Version bump

Supervisor: `0.5.2 → 0.5.3` in all four sources of truth, in lockstep:

• `supervisor/src/index.ts` (VERSION const)
• `supervisor/src/hub-client.ts` (VERSION const — reported in supervisor.hello to the hub)
• `supervisor/tauri/src-tauri/Cargo.toml`
• `supervisor/tauri/src-tauri/tauri.conf.json`
• `supervisor/tauri/ui/package.json`

The next MSI release (`supervisor-v0.5.3` tag, built by `.github/workflows/release-supervisor.yml`) will carry the fix.

Test plan

• Unit tests: `bun test supervisor/test/` — 49 pass / 0 fail.
• Phase 09 canary still passes.
• After MSI build + install: click Launch on a repo, confirm session.start spawns claude and the chat surfaces stream activity in the web UI.

🤖 Generated with Claude Code