Concurrent WebSocket startup can race creating ProjectContext

[rsolmano]

Description

Concurrent WebSocket connections for the same project can both create and start a ProjectContext before either one stores it in _projects. When both contexts initialize the same SQLite spec index at the same time, one can fail with sqlite3.OperationalError: database is locked.

Steps to reproduce

1. Start two WebSocket clients for the same fresh project at nearly the same time.
2. Have both clients call RPC methods, such as spec/list.
3. This showed up in CI in tests/rpc/test_server.py::TestMultiClientIntegration::test_two_clients_both_can_call_rpc.

Expected behavior

Only one ProjectContext.start() should run per project, and concurrent clients should share the same initialized context.

Actual behavior

The second concurrent startup can try to initialize the same SQLite index and fail:

sqlite3.OperationalError: database is locked
app/rpc/server.py:310: await new_ctx.start()
app/rpc/project_context.py:236: needs_rebuild = await self.index.open_and_check(...)
app/spec/index.py:227: await self._conn.execute(line)

Environment

• OS: GitHub Actions Ubuntu runner; local macOS repeated runs did not reproduce
• Python version (python --version): Python 3.11.15 in the failing CI run
• Node version (node --version): N/A
• ThinkRail version / commit: observed while checking PR Generate schemas before Vite startup #8 at 71096ac200c1f240a6e63127a179bca9ecd352d4