mcp: retain spawn configs so reconnect works for ephemeral/managed/builtin servers

[Legoben]

Problem

reconnect_server only looked up installations in locally_installed_servers, so every ephemeral server — managed (backend-proxied), file-based, CLI-run (oz agent run --mcp), and the built-in Factory MCP — failed reconnection permanently with "Installation not found" the moment its transport closed. The spawn consumes the installation and nothing retained it; CLI-ephemeral installs are the only carrier of resolved secrets, so the config couldn't be re-derived either.

Changes

• spawn_configs retention: the manager keeps the exact installation each server instance was spawned with (RetainedSpawnConfig + SpawnProvenance), inserted on every spawn and removed on shutdown/delete. Entries hold resolved secrets/proxy tokens and are never logged.
• Universal reconnect lookup (respawnable_config): consults retained configs first; locally installed servers prefer the live installation (the user may have edited it); the built-in server re-mints its bearer from current session credentials. Gated behind the new McpSelfHeal feature flag — flag off is exactly today's behavior (covered by tests).
• Unflagged reconnect hygiene fixes:
  • the old RunningService is now cancelled during reconnect instead of dropped un-awaited (old stdio children could linger and race the respawn);
  • shutdown_server fails pending reconnect waiters instead of leaking them forever (their spawn was aborted, so the completion callback never ran);
  • authorization_urls is cleared on reconnect (was only cleared on shutdown).

Tests

cargo test -p warp --lib ai::mcp — new in-crate tests cover per-provenance lookup, the live-installation preference, builtin re-mint failure without credentials, waiter drain on shutdown, and flag-off behavior preservation.

Part of the MCP self-healing stack (see stacked PRs above/below in Graphite).

[Legoben]

• mcp: detect dead transports, add Reconnecting state, background reconnect #15540
• mcp: proactively re-mint managed proxy tokens before expiry during agent runs #15539
• mcp: re-mint expired managed proxy tokens on demand instead of failing or misrouting to OAuth #15538
• mcp: broaden reconnect retry, add per-server backoff, keep tools visible during reconnect #15537
• mcp: classify auth failures (HTTP + SSE) and guard OAuth credential deletion #15536
• mcp: retain spawn configs so reconnect works for ephemeral/managed/builtin servers #15535 👈 (View in Graphite)
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.