fix(work): a claimed bench card's solve happens IN THE CARD'S ROOM + typed dispatch struct (#425 slice 1)

[joelteply]

#425 slice 1. A claimed bench card's solve now happens in the card's own room, and the dispatch takes a typed struct instead of a positional list with two bare UUIDs.

Measured first, because the doctrine doc's premise is wrong

Every workspace-trace record carries room_id. Counting nil vs real across the 25 newest trace files:

	turns
roomed	138,088
roomless	13,209 (8.7%)

Per citizen: 90e758b2 8,657 nil / 15,814 roomed — 35% of her cognition happened where no room could see it; fe4dac17 2,133/15,316; e5f4141d 1,177/50,273; a20b3ada 895/489 (more roomless than roomed).

What BENCHMARKS-ARE-ADAPTERS-NOT-A-RUNNER.md overstates: "a detached agent/solve produces no turns, so a citizen can burn 12 acts… and none of it reaches the curriculum." It DOES produce turns. build_workspace_cycle attaches a JsonlWorkspaceCaptureSink for every cycle — live turn, eval fork and solve alike (persona_workspace.rs:551-564); 1,258 trace files, 6.7 GB on disk right now. #427 is the corroboration and I had already accepted it: its whole subject was benchmark forks appending to the SAME traces file as live turns. You cannot contaminate a corpus you do not write to.

So the defect is narrower and still serious: those turns are invisible to the ROOM. No act receipts radiate (apply_act skips a nil room BY DESIGN), so no peer, no human and no ViewState ever sees the work. A citizenship defect, not a curriculum-input one — which is exactly what #425 is named for.

The room was never unknown

Boards are PER-ROOM, so a card's activity is the room whose board holds it — and the wrong-room claim retry directly above already resolved exactly that, inline. Extracted as room_holding_card: one resolver, two consumers, replacing a scan in one place and a shrug in the other.

work/claim now resolves the card's room and dispatches into it. A card no subscribed board holds gets no detached fallback — the claim stands and a work.claim.unplaceable_card probe says why, because inventing invisible work is the thing being removed.

The signature, on two calls from today

"Make sure params are well formed structs and constants, good OOP, not random parameter lists"

The old signature was (ctx, airc, Uuid, WorkCardId, Option<Uuid>) — two bare UUIDs separated only by argument POSITION, so transposing claimer and room compiled cleanly and would have dispatched a citizen's solve under a room's id. Now:

pub(crate) struct StagedSolveDispatch {
    pub claimer: crate::identity::PeerId,
    pub card: airc_work::WorkCardId,
    pub room: airc_core::RoomId,
}

Typed fields make that transposition a compile error, and the next fact this dispatch needs becomes a FIELD, not a sixth argument. ctx + airc stay separate parameters deliberately — they are ambient services, not facts about the dispatch; folding them in turns a value object into a context bag.

"If something is required, remove the option… a required param is caught at compile time not runtime"

room was Option<Uuid> and work/claim passed None — the whole defect, expressed as a type that permitted it. It is now RoomId, so the roomless dispatch is unrepresentable: a caller that cannot name the activity cannot construct the struct. The fallible part is the LOOKUP, which is where fallibility belongs, and it is handled at the boundary before the struct exists.

Adjacent smells named, not silently expanded

• The dispatch roster still carries bare (String, Uuid) tuples (Add claude/snapshot — persistent memory across sessions #396's subject), so benchmark.rs types the id AT this boundary rather than reaching into the roster.
• AgentSolveParams::room stays Option because agent/solve is also operator-invocable; narrowing that param is its own slice, with a comment saying so at the call site.

Validation

cargo check -p continuum-core --features metal,accelerate clean, zero errors.

Live proof owed and NOT claimed: no dispatched claim has run on this build. The falsifiable prediction is that a claim-fired solve's acts now radiate receipts into the card's room, and that the nil-room share of new traces falls from the 8.7% measured above.

Still open on #425 (this is slice 1, not the card)

• ExperienceRecord::from_lived_turn has zero production callers tree-wide — the experience stream is fed by from_kanban_grade alone, so no lived turn (bench or chat) ever becomes an experience record. Wider than Gap analysis: 49 open issues, 20 closed #425.
• persona::recorder::record_turn is called only from persona/response.rs (the respond() path); cognition/ never calls the recorder, so a solve writes workspace traces but no persona-respond fixture.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo