fix: harden org-scoped MCP config chat gating, updates, and visibility - #28065

updateMCPServerConfig merged request fields onto the param middleware's
snapshot, taken before the request body was parsed, so an overlapping
update could have its changes to omitted fields silently reverted.
Re-fetch the row on the transaction handle before constructing the
merged update, restoring the pre-existing behavior.

…er refetch errors

A failed background refetch flips the MCP query's isSuccess off while
TanStack Query v5 retains the cached data, so gating the composer on
isSuccess blocked sends despite a usable organization MCP list. Gate on
missing data instead; the error alert still surfaces refetch failures.

The auditor roles' MCP config read grant exists so auditors can inspect
the resources audit logs reflect, but the full-view gate only checked
update permission, leaving auditors with the member view: redacted
management fields and 404s for disabled configs. Before org scoping the
site auditor's deployment-config read selected the full view. Gate the
full view on org audit-log read as well, which site and org auditors
hold and ordinary members do not.

The cross-org concealment helper prefixed every case with
/mcp-servers, so the callback case hit a router 404 without invoking
the handler; the callback lives on the frozen /mcp/servers route.
Disconnect deliberately returns 200 for callers without a token so
config IDs cannot be probed, which the 404 expectation contradicted.
Also pin that legacy MCP selection migration is a no-op while the
server list has not resolved.

…ect body concealment

Add a cross-organization callback test proving tokens bind to
(user, org-scoped config) with no cross-org reuse, and restore the
full-response equality check that a concealed config's disconnect is
indistinguishable from a nonexistent ID.