orch-frontier

name	orch-frontier
description	Execute a ticket dependency graph by rolling frontier dispatch — every ready ticket in flight. Use whenever items have dependency edges.
role	none

Require: one finite acyclic run ticket graph issued through tickets.py, plus its bound from the root ticket or caller.

Dispatch every ticket whose depends_on are complete, one lane each. Claim through tickets.py claim; take its packet (refusal is a cut defect); spawn one fresh child under delegation §1–§2, role per roles §4, host depth per profiles, applying its executor. Each ticket takes one outside-independence path: only for unchecked independence: checker with an authored-here oracle, dispatch the same claim's --executor orch-critique checker packet, and never for gate-deferred, already checked, or pre-existing-only tickets. Then, where its pass invalidates a judged oracle, dispatch that packet's --executor orch-verify form; otherwise re-run invalidated deterministic oracles at the checked identity here (verification §10).

For v2, readiness, claim, and packet require that root_generation, cut_generation, and assignment_seal resolve to one sealed run-state snapshot and validation receipt. Refuse draft, merely validated, stale, missing, or mismatched generations; never substitute the latest. After an accepted tickets.py amendment-request disposition, recompute the frontier: only what tickets.py seal newly sealed proceeds. The absence of v2 fields means v1. A root cut reader is the exception: take it with three or more <id>.NN or after a cutcheck advisory; units stay pending until checked_by and this engine's cutcheck.py re-run, the re-verification, exits 0. Below that threshold cutcheck accepts the cut alone. Accept every return once through orch-integrate; suspended parks it, any other grades isolation and integrates per the pack, conflicts through its binding. After each merge batch run the standards owner's required checks on the integrated tip, the run's notes carrying the tip's revision. A lane runs its ticket's own oracles, nothing wider; its green provisional until the tip's, and a red tip blocks the next dispatch but its repair's.

Watch per profiles; recompute on results, tickets, suspension, or stale claims. Re-checks run bound-check: overdue without motion since the bound parks suspended through the join-side status path, Handoff naming bound, last_motion_at; overdue with motion reports over-bound. Promote via tickets.py ready, dispatch at once, report skipped; name terminal dependencies blocking pending tickets. Parked claims remain live, dependents wait, unsatisfied exclusions exit with the remainder. At quiescence read durable successors.md. With a planned entry and the run complete, return the successor trigger plus the predecessor's accepted ## Result identity to the plan's materialization owner, who materializes it and replaces the plan before completion is reported (work-item). Otherwise exit; limited when the bound leaves tickets open.

Never: hold a ready ticket back to batch it; hide a blocked subtree in a summary of successes; re-order the graph to dodge a failure.

Return: status; per-ticket results by identity; the graph's terminal state as verification; successor trigger and materialized run when successors.md requires one; the open remainder with what blocks it.