[kerdofficial]

status returned every operation, activity and instance ever recorded, and instances were never deleted. On a real machine the MCP vectis_status result reached 1.5 million characters (the result was also sent twice), remote status failed past 256 KiB, and operation/wait fetched the whole snapshot every 200 ms to find one record.

• The snapshot holds every unfinished record plus the newest 100 finished operations and activities and the newest 20 stopped instances. Stopped instance records older than a day beyond that are deleted.
• New GET /v1/operations/:id and /v1/activities/:id endpoints back operation, wait and activity lookups locally and remotely, and service-internal lookups read the store directly so old records stay reachable.
• MCP vectis_status returns a compact summary and points to the operation and activity tools for detail; results are no longer duplicated.

Stacked on #106.

Proof: store, HTTP, client, remote and MCP tests, including an MCP status under 64 KiB with 2000 operations; on the dev deployment the bounded service ran real jobs and an operation outside the snapshot was still read directly. runner-access.test.ts pwsh hook tests time out only under parallel load and pass alone, as on main.

[kerdofficial]

claude-opus-5-5 RESPONDING ON BEHALF OF DANI

Superseded by #109, which combines this change with the rest of the night's reliability work in one PR.