refactor(server): move the MCP handlers' SQL into repositories - #50

What moved

The MCP handlers (mcp_servers.rs, mcp_oauth.rs, mcp_oauth/shared.rs, mcp_oauth/wizard.rs, mcp_store.rs, mcp_tools.rs) no longer carry SQL. All 50 sqlx::query* calls move into four repositories under crates/server/src/services/, with the statement text byte-for-byte unchanged (checked mechanically against the old handlers), the same binds and the same fetch kind:

• mcp_server_repository — mcp_servers reads/writes; the update transaction with its credential purge, the delete / bulk-delete transactions with the install-count decrement; last_error updates; the unique-violation → 409 mapping (moved from the handler).
• mcp_credential_repository — per-user and shared credentials: the upsert (advisory lock + first-is-default), revoke-and-promote and set-default transactions, shared status/token reads, inserts used by server create.
• mcp_store_repository — templates, installs, the registry-sync upsert and prune, the install advisory lock (the lock-key constant moves with it).
• mcp_tool_repository — the tool catalog count and page; McpToolRow moves here (openapi import updated).

create_server and sync_registry interleave non-SQL work inside their transaction, so they keep the transaction and pass &mut PgConnection into the repository. Every other transaction now lives in the repository. Local FromRow row types moved alongside their queries. insert_shared_credential_from_wizard and the upsert_credential / upsert_shared_credential handler helpers are gone; callers use the repository directly. Permission checks, validation, encryption, registry/pool sync, cache invalidation and audit stay in the handlers, in the same order.

No behaviour change

New crates/test-support/tests/admin_mcp_catalog.rs (13 tests) covers what the MCP suite didn't reach: server get/patch/delete incl. 404s, name/prefix 409s on create and update, unknown template 404, active-only tool counts and list order, shared_static_token at create, background last_error set/clear (create and shared-token paths), shared-credential status/paste/revoke, shared authorize preconditions, connection listing filters, set-default / revoke / 404s, template by slug, the registry sync (upsert, update, prune, installed kept, invalid template skipped, auth-shape derivation), categories, and the tool catalog's search / server filter / paging / include_user_tools.

The new file passes on the code before this change (handlers stashed) and after it: 13/13 both times.

Found while writing them, not fixed here (verbatim SQL): revoking a user's default connection returns 500 — the promote statement selects id from mcp_user_credentials, which has no id column. Follow-up PR.

Checks run locally

• cargo fmt --all --check
• cargo clippy --workspace --all-targets -- -D warnings
• cargo clippy --workspace --lib -- -D warnings
• cargo nextest run --workspace --lib --bins --tests — 687 passed
• full integration suite (--run-ignored only --profile ci) on own containers: all MCP tests pass; the 37 ClickHouse-dependent failures in the first run were my ClickHouse container being OOM-killed mid-run, and all 37 pass on rerun after restarting it.

🤖 Generated with Claude Code