queen: a review with no lane free waits for one in a bounded queue - review_lanes.t27, behind TRIOS_QUEEN_REVIEW_LANES=semaphore #1737

Part of #1712. Night loop ledger #1729, item 4.

Problem

With scarce lanes the adaptive reviewer pool (#1715, #1716, reviewer_sizing.t27) loses to the fixed 4 on the benchmark's second table. On the same seeded input with 3 free lanes:

• overload: 360/660 against 429/660;
• burst: 352/365 against 365/365.

The cause is in how a review uses its lane. It holds the lane only during its model call, and it asks for one about 30% of the way in. If no lane is free at that point, it answers wait. A pool capped at the lane count therefore leaves lanes idle during the first 30% of every review. The fixed 4 overlaps that phase, but at the price of hundreds of lane-less wait visits. Each of those visits throws away a witness and a criteria measurement, and due_again pushes the row back.

Change

1. The spec is gHashTag/t27 specs/queen/review_lanes.t27 (spec(queen): review_lanes.t27 - a review with no lane free waits for one in a bounded queue t27#8579, PR spec(queen): review_lanes.t27 - a review with no lane free waits for one, in a bounded queue (Closes #8579) t27#8582). A review that reaches its model call with no lane free waits in a bounded queue. One actor owns the queue. The rules: oldest first; one waiter per review in a model call; the wait ends once the 120 s call no longer fits the 300 s row. The pool may exceed its lanes by the measured share of a review spent without a lane.
2. Runtime: a review-lanes actor. It is the first child of the reviewer domain and owns the queue. Reviews reach it through a LaneGate, which waits by call with the card's timeout. The gate is wired into the reviewer's model-call step in queen-tick.ts. All of this sits behind TRIOS_QUEEN_REVIEW_LANES=semaphore, off by default. It works with and without TRIOS_QUEEN_REVIEWER_ADAPTIVE and TRIOS_QUEEN_TURN_STOP: a stopped turn leaves the queue or releases its lane.
3. Benchmark: the second table of tests/api/queen-actors-bench.test.ts runs the same draws for four variants: fixed 4, adaptive, fixed 4 + queue, and adaptive + queue.

Acceptance

• Adaptive + queue at least ties the fixed 4 where adaptive lost on lanes, and does not lose where adaptive won.
• No new failure in the api suite.
• The numbers are posted on Queen moves to the actor model: the rule in every AGENTS.md, an MVP, a benchmark, then the swap t27#7851.