0.7.4

Added

• The issue-flow sweep flags a stalled chain head: a blocker chain three or
  more issues deep, counting the head, whose claimed head has a red open PR.
  Comment-only, on the head alone, naming the remedy — service the red, or
  re-order the chain (#440, #426).
• lib/checks.sh — the rollup classifier both reconcilers now source, so the
  tripwire grades a head with the same instrument that decides
  blocker:ci-red and the two can never disagree (#440, #136, #139, #208).
• Flag idle queues, deeply serial blocker chains, and dependency cycles from the issue-flow sweep's existing board snapshot (#426).
• Triage now puts measured flow-wide breakage at the front of a collision
  chain while ordinary work keeps arrival order (#425).
• Front-row insertion records its evidence and edge direction, preserves a
  claimed issue's labels, and orders overlapping work at merge (#425).
• actions/ci-rerun: a base-repo workflow that services rerun-owed by
  starting the rerun a fork PR's author cannot. Its job holds actions: write
  for one run, checks out nothing from the head, and runs no PR-authored code
  (#424).
• Four gates, all measured at service time and none inherited from the label:
  the actor is a fleet identity in .github/labels.conf, the head still is the
  one the evidence names, the run concluded failure, and it is on attempt 1
  (#424).
• A started rerun removes rerun-owed and comments the new attempt's URL; a
  refusal leaves the label standing and comments which gate refused and what a
  human would have to do. There is no schedule and no retry — the workflow acts
  on one label event and stops (#424).
• Consumers get the servicing at their next pin bump: docs/CONSUMERS.md
  carries the caller stub and the permissions block, without which the
  workflow refuses every time (#424).
• Only a run created before the evidence comment that names the head is a
  rerun candidate: a pull_request_target run carries the head's SHA, so the
  workflows the label event itself wakes are in that list too and must never
  be rerun in the red run's place (#424).
• rerun-owed: a PR-only label for a head that is red on a rerun no agent may
  start. The builder sets it with evidence opening 🔁 rerun owed at head <sha>; the reconciler clears it when that head's checks leave failing, or
  when the named head is no longer the head (#423).
• While rerun-owed stands at a head, blocker:ci-red is not asserted there:
  that label says the builder owes a fix, and on a fork PR's base-repo run the
  builder owes nothing. It is not a blocker and never enters the set (#423).
• A sixth park shape: a red head whose rerun is a right the builder does not
  hold parks the claim and frees the build slot. Every other red head is still
  the builder's, and the ci-red wake skips the labelled one (#423).
• rerun-owed is not exempt from the 48-hour staleness clock: a head standing
  under it still goes stale, and that stale asks for a poke of whoever
  services the rerun, never a fix from the builder. A flag earns an exemption
  only by carrying an escalation clock of its own (#423).

Changed

• Document the issue-flow sweep's comment-only idle, deep-chain and cycle
  tripwires, including placement, shape-keyed deduplication and shared
  threshold ownership (#441).
• The ci-rerun roster precondition in docs/CONSUMERS.md drops its census
  of the governed repositories' .github/labels.conf files, which was false:
  identities that build as well as review are named in the roster the gate
  reads (#437).
• In its place the paragraph gives a directive the reader runs against their
  own conf — the three fields gate 1 reads — rather than a measurement of
  other repositories that nothing re-checks when a roster changes (#437).
• The one-build-at-a-time slot is stated as the builder's own, across every
  repository they work in: the self-check is whether an unparked claim is
  held anywhere, never whether the board in front of you is clean (#430).
• The slot sentence stops carrying a second test for when a build ends. A
  claim holds the slot until it is parked or released, so the park list is
  the only definition of finishing and the two can no longer disagree (#430).
• The claim comment now asserts the slot: no unparked claim in any
  repository, naming any parked claims held elsewhere with their shape
  (#430).
• AGENTS.md's router row now sends the builder to an ordered chain of PRs,
  normally one — the fourth surface of the singular arithmetic, and the one
  the mint's regex could not see behind its backticked ready (#420, #429).
• One issue now means an ordered chain of PRs, normally one. A PR carries at
  most five rounds; at the close of round 5 the builder continues the branch
  in a successor PR and the predecessor closes as the ledger (#420).
• The cap is a consensus-surface rule and not a byte defence: the longer a PR
  runs, the harder it is to bring the whole panel onto one head (#417, #420).
• A cut spends every approval. The review target becomes the successor, no
  verdict is owed on the closed predecessor, and rounds are numbered per PR
  so the successor's first round is round 1 (#420).
• The issue gains a triage-maintained ## Pull requests list in chain order,
  present only once a chain exists and triggered by the builder's cut
  comment; it is the only place a chain stays navigable (#420).
• Triage now sizes accepted work for bounded review rounds and splits oversized work into ideally disjoint epic children, treating later round growth as evidence for the next mint (#419).
• The PR body's ## Round log is a rolling summary rather than a verbatim
  mirror of each round's reply: ### Current state is rewritten every
  round, ### Rounds gains one row of facts plus two prose cells, and
  handoff waits on those cells being filled (#418).
• A round's reply shrinks to a short comment naming the round and the head,
  so the comment thread stops growing with the body it used to duplicate
  (#418).