EPIC: best-in-industry dispatch -- why 32 of 70 bees idle with a full backlog, and the roadmap to fix it #9976

The owner asked for a roadmap that makes the Queen's dispatch the best in the industry, and for the reason the bees stand idle while the backlog is full. This epic answers both. It is a tracking issue: it has no Boundary on purpose and is never dispatched itself.

Where we stand (measured 2026-10-17 14:37Z, /queen/status)

Why the bees stand idle: the work exists, but it does not reach the dispatcher as eligible issues.

1. Hot files. A few foreign files (bootstrap/src/compiler.rs, bootstrap/src/main.rs, scripts/tri_loop/t27b.py, cli/tri/src/main.rs) sit in the Boundary of dozens of issues, so one held issue shuts out the rest: fileConflict 166.
2. The roadmap is not in the queue. Epics are not split into Queen-shaped sub-issues, so they count as missingBoundary (79), and the planner does not read the roadmap (Queen roadmap from one .t27 source, live from the epics, read by the planner; plus self-diagnosis rules #9022).
3. Done work stays open. 508 issues are complete but open, and are re-skipped every tick.
4. Master is red on duplicate-bodies, cli-tri and t27b-native-ratchet, so no bee PR lands (duplicate-bodies: share tree helpers across eight gen_w bench ports #9633, cli-tri: retire deleted-workflow test and refresh measured census pins #9634 / PR policy: record approval for the 19-line gates test deletion (Refs #9634) #9639, t27b-native ratchet is red on every master run since it landed: ledger 099ac2224 predates 16 improvements and 40+ new specs #9600 / PR t27b: repair crypto and rust_convert; distinguish PR ledger lag #9637).
5. Contracts are not wired. Most fixes already exist as specs/queen/*.t27 cards, but the runtime (gHashTag/trios) does not run them yet, and hand-written TypeScript needs the owner label.

Phase 0 -- unstarve the swarm (days)

• Master green: duplicate-bodies: share tree helpers across eight gen_w bench ports #9633 (duplicate-bodies, pure .t27, in progress), PR policy: record approval for the 19-line gates test deletion (Refs #9634) #9639 (cli-tri, owner label), PR t27b: repair crypto and rust_convert; distinguish PR ledger lag #9637 (t27b ratchet, owner label)
• Give the missingBoundary issues a Boundary or mark them as epics (in progress)
• queen: reconcile issue state from PR linkage -- completed-but-open issues close one at a time, with evidence #9968 reconcile: close completed issues one at a time, with the PR as evidence
• queen: a hot file is named and split -- a path that shuts out more than N issues becomes a finding with a split plan #9975 hot files: name them, split them, and stop holding paths that bees may not edit by hand
• t27b ts: emit the Queen's rules as TypeScript with function bodies, from .t27 (supersedes the blocked gen-ts PR #7872) #9994 t27b ts: the bridge. It emits the cards as TypeScript with function bodies, from .t27 (ts_emit.t27 / ts_plan.t27, the shape of t27b verilog). It supersedes gen-ts: lower the bodies of pure functions, so a host imports a spec's rule instead of copying it #7871 / PR gen-ts: lower the bodies of pure functions #7872, which the own-language gate blocks for good (441 Rust lines against a 40/80 limit that no label lifts).

Phase 1 -- flow (Kanban, Little's law, merge queues)

• queen: service levels with an error budget -- written targets, and dispatch slows when review falls behind #9974 SLOs and error budget; dispatch follows review capacity
• queen: event-driven requeue -- a skipped issue wakes when its blocker releases, not by rescanning 826 issues every tick #9967 event-driven requeue (Kubernetes QueueingHint pattern)
• queen: a merge queue that batches green PRs and bisects a red batch, so landing keeps up with finishing #9973 merge queue: batch and bisect (Uber SubmitQueue, GitHub merge queue, Google TAP)
• Queen wiring: one chooser call per wake fills every free slot (offers_per_wake, resident queend) trios#1710 / feat: extend IGLA RACE cross-target to systolic PE + gen-backend findings #1881: one chooser call fills every free lane (TRIOS_QUEEN_DISPATCH_BATCH, merged, flag state not visible)
• Queen wiring: reviewer concurrency from the spec instead of a fixed 4 trios#1711: reviewer concurrency from specs/queen/reviewer_sizing.t27
• Queen wiring: file owner actor admits disjoint edits instead of the fileConflict skip trios#1709: a file-owner actor admits disjoint edits instead of the fileConflict skip
• queen: wire conflict_step into the publisher (update-branch / redo conflicting bee PRs) #7573: publisher conflict step (update-branch, redo conflicting bee PRs)

Phase 2 -- reliability (Temporal, BEAM)

• queen: bee key failover -- on 429/1305 the next attempt goes to another key, then the other pool, instead of retrying the same key until the bee dies #9993 key failover: on 429/1302/1305 a bee moves to the next key, then the other pool (17 of 93 bee turns died after 0 steps on 2026-10-17)
• Queen wiring: record the exit reason of every dispatch in the activity event and run the spec's conformance rows trios#1708: record every dispatch's exit reason (specs/queen/dispatch_exit.t27, 24 conformance rows)
• queen: a runner bee that committed nothing is an empty attempt after the grace, not a 6-hour wait (reviewer.t27 no_bundle_is_empty, behind TRIOS_QUEEN_NO_BUNDLE_EMPTY) trios#1886: a bee that committed nothing is an empty attempt after the grace, not a 6-hour wait
• Queen: a redispatch after a spent ceiling should get a full set of send-backs (option (a) of #1769, depends on t27#8768) trios#1806 + queen: preserve bee worktrees and bound retries with a standalone policy card #9552: a retry policy instead of the retry-ceiling lock (tools/queen/refile.py re-files today)
• Queen dispatches issues that already have an open PR closing them (4 duplicates on 2026-10-10) trios#1736: never dispatch an issue that already has an open PR closing it
• Queen deploy drain almost stops dispatch for ~30 min (rounds gated during drain) trios#1705: a deploy drain does not stop dispatch for 30 minutes
• queen: review rounds fail to read the diff of finished work (1105 lines in 46 min) #6116: the review round reads the diff of finished work

Phase 3 -- intelligence (where agent schedulers are going)

• queen: order work by value -- cost of delay over size (WSJF) and the number of issues it unblocks #9971 value: WSJF plus how many issues a task unblocks (critical path)
• queen: route each task to the cheapest tier that can finish it, and escalate one tier on send-back #9972 routing: the cheapest tier that can finish the task, one tier up per send-back (FrugalGPT, RouteLLM)
• queen: idle lanes run speculative attempts -- a second bee on the highest-value long-running issue instead of standing idle #9969 speculative attempts on idle lanes: best-of-N and backup tasks for stragglers (MapReduce, Codex, Devin)
• Queen roadmap from one .t27 source, live from the epics, read by the planner; plus self-diagnosis rules #9022: the roadmap as one .t27 source the planner reads

How a card lands

Each new card above is Queen-shaped, with a Boundary of one new spec file and its seal, so the cards never conflict with each other and bees can take all eight at once. The bar is specs/queen/dispatch.t27 in PR #9961: 0 surviving mutants, 0 vacuous tests. Wiring follows through t27b ts (#9994), not hand-written TypeScript.

Refs #6657 #9021 #9891 #6595 #9025