queen(actors): the reviewer's worker pool follows its backlog and its lanes - reviewer_sizing.t27, behind TRIOS_QUEEN_REVIEWER_ADAPTIVE #1715

Part of #1712 (item 3: reviewer concurrency read from the measured backlog).

The actor reviewer starts a fixed REVIEWER_CONCURRENCY = 4 workers (queen-review-actors.ts). gHashTag/t27#8212 already says how many rows may be reviewed at once: reviewer_concurrency(keys, lanes_per_key, free_mb, mb_per_review, waiting), capped at 16. gHashTag/t27#8273 adds reviewer_sizing.t27, which covers what the runtime feeds that rule and how it grows and shrinks the pool.

What

• The reviewer domain sizes its worker pool from the card, behind TRIOS_QUEEN_REVIEWER_ADAPTIVE=1 (default off).
  • It re-reads the lanes and memory on every wake.
  • It grows by starting transient workers under a dynamic supervisor (start_answer).
  • It shrinks by stopping idle workers once worker_retires says so.
• The lanes are counted by the review sweep's own arithmetic. Each key is capped at what its provider is measured to carry: two requests a key on z.ai (key_free_lanes).
• Tests: the pool follows the card for backlogs of 0, small and huge, and never exceeds the lanes.
• Benchmark: fixed 4 against adaptive on the same seeded input as queen-actors-bench.test.ts.
  • Cases: quiet, steady, overload at 120/h, and a burst.
  • Measures: throughput/h, p95 wait, the most provider calls at once, and backlog drain time.

Boundary

• trios/agent-server/specs/queen/reviewer_sizing.t27, .wasm, specs/PIN
• apps/server/src/api/services/queen-review-actors.ts, queen-actors-dynamic.ts (new), queen-reviewer-sizing-card.gen.js (new), queen-review-loop.ts, queen-tick.ts (the capacity reader and the start call)
• apps/server/tests/api/queen-review-adaptive.test.ts (new), queen-actors-bench.test.ts

Acceptance criteria

• bun test tests/api/queen-actors*.test.ts tests/api/queen-review-adaptive.test.ts passes.
• The api suite fails no test that actors-next does not already fail.
• The original benchmark table is unchanged: the fixed path behaves byte for byte as before.
• The benchmark numbers are posted on the PR and on Queen moves to the actor model: the rule in every AGENTS.md, an MVP, a benchmark, then the swap t27#7851.