ci: stabilize nightly Windows Playwright shards (#14726)

Every nightly since Aug 20 lost 1-2 Windows shards on attempt 1:
- backend boot killed by the 90s webServer timeout mid-migration
- 20-30s mid-test API stall while a cold uvx cache resolves mcp-proxy
  for the auto-registered lf-starter_project MCP server
- Linux-calibrated timeouts (5s expect, 20s action, raw 10s waits) and
  two racy assertions (lock-flow edge delete, outdated-actions count)

Windows-conditional timeout headroom in playwright.config.ts, a uvx
mcp-proxy pre-warm step in the Windows shard setup, TIMEOUTS-convention
waits in the affected specs, and a selection-verified edge delete in
lock-flow. Linux behavior is unchanged.