Enforce that every attached server contributes a briefing

[lsm]

The contract test. Moved out of #4780 and to the end of the ladder, because it was misplaced: the invariant cannot bite until servers actually contribute, so writing it at Build time would have meant a test asserting something nothing yet does.

every attached MCP server contributes a briefing

Same spirit as check:operation-names. This is what turns #4772 from something exploratory QA discovers into a CI failure.

Depends on #4791 (third-party server decision) — that determines whether the invariant covers app-registry servers, is narrowed to built-ins, or is stated over generated contributions.

Rung 5c of #4778. See ADR 0007.

[lsm]

Blocker for how this test can be written, found by the pins in #4806.

The obvious binding surface is dead. requiredServers is [] for every role in resolveSpaceMcpSessionPolicy, missingMcpServers has no production caller, and the three onMissing*McpServers callbacks are assigned in five places and invoked nowhere — query-runner.ts only logs an MCP snapshot. So "required servers" is unenforceable today for every session kind, and a contract test bound to it would assert nothing.

Two options: revive that mechanism as part of this rung, or bind the test to the effective set from getEffectiveMcpServers() instead, which is what clause (a) of the amended invariant in ADR 0007 actually describes and what query-options-builder.test.ts already asserts over.

I lean to the second — the effective set is the thing the model really receives, and requiredServers being empty everywhere suggests it was a seam that never found its use.

Also note the invariant needs a second qualification beyond the third-party one: agent-memory and db-query are not attached uniformly to Space sessions (#4808 for workers, #4807 for direct task workers), so clause (b) cannot assume every Space session has the same built-in set.

[lsm]

Blocked on a type change, not just on #4791. Clause (b) of the invariant distinguishes authored from self-describing contributions, but CapabilityContribution as merged in #4780/#4804 is the flat { server, briefing } — there is no kind discriminator to assert on. #4783 now carries introducing it, along with widening McpServerAttachment.config so an SDK server is expressible at the type level.

Until both land, this test can only assert clause (a), the two-way accounting between queryOptions.mcpServers and the session's contributions. That is still worth having and is the part that would have caught #4772, so it may be worth splitting: land clause (a) now, clauses (b) and (c) after #4783.

Reminder of the other blocker already noted above: bind to the effective set from getEffectiveMcpServers(), not to requiredServers, which is [] for every role.

[lsm]

Third blocker, from #4844 (which lands the union this test needs).

Clause (a) accounting cannot pass yet for two session kinds. Space chat and workflow workers hold built-in MCP servers but assemble no briefing at all — there is no briefing path for those kinds. So the two-way equality between queryOptions.mcpServers and the session's contributions fails for them, not because a contribution is missing but because nothing assembles one.

Giving them a briefing path is a scope question (#4795 for space chat, #4782 for the worker kinds), not a capability one, so it belongs to the other lane and has to land first.

Running total of what gates this rung:

1. The kind discriminator — arriving in feat(briefings): agent-memory and db-query contribute authored briefings #4844.
2. requiredServers is [] for every role and the onMissing* callbacks are invoked nowhere, so bind to getEffectiveMcpServers() instead.
3. A briefing path for space chat and workflow workers, or an explicit decision that those kinds are exempt from clause (a).

Point 3 is the one worth deciding early, because "exempt" would be a hole in exactly the shape #4772 came through.

[lsm]

Parked under #4877 until the daemon is dogfoodable.

The current rule is no MCP policy or gates except for correctness — the effort is the shortest path to running the dev-tip daemon as a daily tool. This is not rejected and nothing here needs redoing; the analysis stands and gets picked up with the rest of the parked set.