queen: wire keyed_batch.wasm so the bee admission makes one queend choose per batch, not per issue #2043

Problem

On 2026-10-17 the keyed-actors dispatch trial (TRIOS_QUEEN_DISPATCH=actors, build 42042a5) started 2 runners in 20 minutes. The round loop starts 18.5 in the same time, so the trial was reverted (evidence).

The cause is serial admission:

• handRoundToBeeActors (queen-bee-actors-round.ts:216) readies about 826 candidates.
• The single bee-admission actor then answers them one at a time with await deps.admit(ask.issue) (queen-bee-actors.ts:298).
• Each of those is one queend choose call for one issue (round.choose(issue, ...) at queen-bee-actors-round.ts:152, which becomes kind:'choose', candidates:[issue] at queen-tick.ts:2267).
• The loop asks queend once per round about the whole list (queen-batch-choose.ts).

Spec

gHashTag/t27#10051 adds specs/queen/keyed_batch.t27 (closes gHashTag/t27#10033). That card is the policy. This issue covers only the thin wiring that reads it, in the same way keyed_guard.wasm is wired today. No policy lives in TypeScript.

Wiring

1. Vendor the card.
   • Copy keyed_batch.t27 into agent-server/specs/queen/.
   • Build queen/keyed_batch.wasm the way the other cards are built: t27c gen-c followed by zig 0.16.0 cc --target=wasm32-wasi -Oz -DNDEBUG -mexec-model=reactor, exporting the card's functions plus lane_free, admit_answer, wake_count and batch_quota.
   • The reproducible lab sha256 is 1ed6bcb13de9ff4c462ff6d0f8a1ba6e3c590d0235756f3aa4ce4bd8fcacfd8a (1187 bytes).
   • Pin both files in specs/PIN.
   • Load the wasm with loadCardWasm('queen/keyed_batch.wasm'), next to the guard card (queen-bee-actors.ts:112-118).
2. Give the admission actor an open batch (group commit).
   • On each ready, ask the card joins_batch / answer_now. Asks that do not join are answered at once, as today.
   • When batch_goes_out is true (no choose in flight), take batch_take(open) asks and send them as ONE choose. Asks that arrive meanwhile wait for the next batch. There is no timer.
3. Make deps.admit batch-shaped.
   • It takes a list of issues and returns one word per issue: PICK, LATER or NO (the card's word_of).
   • Behind it, make one queend choose with all the candidates, reusing readAnswer / planBatch from queen-batch-choose.ts.
   • Cap new lanes with the card's new_lane_quota. Kept asks (a restart on a lane the issue already holds) sit outside the quota.
4. Answer each ask in order with batch_answer(kept, holds_lane, free, word), re-reading lane_free after every grant. batch_grants is the cross-check: a batch grants exactly its quota.
5. Wake: replace the wake_count call (queen-bee-actors.ts:254) with batch_wake_count. A freed lane wakes every waiting issue into one batch instead of one issue per lane.
6. Bound the calls.
   • askQueend (around queen-tick.ts:1330) spawns queend with no timeout. Give it the card's CHOOSE_BOUND_MS (12000).
   • Keep admitAfterMs (queen-bee-actors.ts:226, 30000 today) above twice that bound. The card's ask_bound_ok must hold.
7. Treat a failed choose as a wait, not a refusal. A choose that throws or times out must become LATER via word_after_choose. Today .catch(() => false) turns it into a refusal.
8. Measure.
   • Expose the dispatcher's counters in the Queen actors measured log line: chooses per round, asks per batch, grants and waits.
   • Re-run the trial against the loop and post the before/after runner starts per 20 minutes on Queen moves to the actor model: the rule in every AGENTS.md, an MVP, a benchmark, then the swap t27#7851.

Done when

• The vendored card and wasm are pinned.
• The admission makes one queend choose per batch, not one per issue.
• A trial run's runner starts per 20 minutes are posted on Queen moves to the actor model: the rule in every AGENTS.md, an MVP, a benchmark, then the swap t27#7851 next to the loop's 18.5.