[Fix] Tasks can mistake unavailable integrations for unconfigured ones

[roomote-roomote]

​Opened on behalf of @mrubens. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Track failures while loading user/catalog and deployment custom-stdio MCP configuration during task startup.
• Keep startup best-effort, but add a secret-free developer instruction that prevents the running agent from treating missing tools as proof that an integration was never configured.
• Cover both failure paths and verify internal fetch errors do not enter agent instructions.

Why this change was made

Roomote intentionally continues task startup when integration configuration cannot be fetched. Previously those failures existed only in worker logs, so the running agent could incorrectly tell users that a temporarily unavailable integration was not configured.

Impact

Tasks still start during MCP configuration outages, while agents can now explain that integration availability could not be verified and direct users to retry or ask an administrator to inspect Integrations settings.