<?xml version="1.0" encoding="utf-8"?>
<entry xmlns="http://www.w3.org/2005/Atom" xml:base="https://docs.spec-kitty.ai/adr/3.x/2026-09-15-1-ci-main-verdict-topology/">
  <id>https://docs.spec-kitty.ai/adr/3.x/2026-09-15-1-ci-main-verdict-topology/</id>
  <title>ADR: Main-Tip Verdict Topology — Adjudicating the #4347/#4371/#4360-A/#4430 Coupled Cluster</title>
  <updated>2026-09-15T00:00:00.000Z</updated>
  <author><name>Spec Kitty</name></author>
  <link rel="alternate" type="text/html" href="https://docs.spec-kitty.ai/adr/3.x/2026-09-15-1-ci-main-verdict-topology/"/>
  <summary>Ratified coupled levers (main-concurrency, fan-out cap, main-green source, terminal-cancel verdict) so a landed main tip proves itself green honestly, in bounded time.</summary>
  <content type="html">&lt;h2 id=&quot;context-and-problem-statement&quot;&gt;Context and Problem Statement&lt;/h2&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Ratified by the operator (agentic-framework-core-team) on 2026-09-15.&lt;/strong&gt; The
recommended coherent bundle below — &lt;strong&gt;1a / 2a / 3a (+3b target) / 4a (+4b backstop)&lt;/strong&gt; —
is accepted; a follow-on implementation mission stages it per the rollout plan. Every
mechanism claim carries a &lt;code dir=&quot;auto&quot;&gt;file:line&lt;/code&gt; or &lt;code dir=&quot;auto&quot;&gt;issue#&lt;/code&gt; (verified at the pre-mission HEAD
&lt;code dir=&quot;auto&quot;&gt;e72e13ca01&lt;/code&gt;; the implementing mission re-verifies exact lines, as &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; has since
advanced and #4360-B landed); hypotheses are labelled &lt;strong&gt;[HYP]&lt;/strong&gt;.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Epic &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4437&quot;&gt;#4437&lt;/a&gt; (“CI pipeline
honesty”, P1, milestone 4.0.0) rests on one invariant, set by ADR
&lt;a href=&quot;/adr/3.x/2026-07-17-1-red-main-is-honest-ci-is-release-authority/&quot;&gt;&lt;code dir=&quot;auto&quot;&gt;2026-07-17-1-red-main-is-honest-ci-is-release-authority&lt;/code&gt;&lt;/a&gt;:
mainline CI status is the release authority, which only holds if &lt;strong&gt;red means a real
regression and green means actually proven&lt;/strong&gt;. This ADR addresses the sub-cluster of
#4437 where that invariant is broken not by a single bug but by a &lt;strong&gt;shared topology&lt;/strong&gt;:&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Shared root.&lt;/strong&gt; Main-branch CI gates are built on PR-head-triggered &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt;
fan-out, and a non-green terminal state has no honest way to report itself.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;p&gt;Four issues are the same shape expressed as four symptoms. The epic body (and #4371
itself) names #4347/#4360/#4371 as “one remediation, solve together”; #4430 rides the
same fleet-verdict surface and is folded in here for the same reason. &lt;strong&gt;They are one
ADR, not four, because a point fix on any one lever reopens the others&lt;/strong&gt; (#4371:
“A point fix on any one of the three leaves the other two … Resolve together”).&lt;/p&gt;
&lt;p&gt;The four in-scope mechanisms, all confirmed against the tree at HEAD &lt;code dir=&quot;auto&quot;&gt;e72e13ca01&lt;/code&gt;:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;#4347 — main-concurrency cancel-cascade.&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;.github/workflows/ci-router.yml:35-37&lt;/code&gt;:&lt;/p&gt;
&lt;div class=&quot;expressive-code&quot;&gt;&lt;figure class=&quot;frame not-content&quot;&gt;&lt;figcaption class=&quot;header&quot;&gt;&lt;/figcaption&gt;&lt;pre data-language=&quot;plaintext&quot;&gt;&lt;code&gt;&lt;div class=&quot;ec-line&quot;&gt;&lt;div class=&quot;code&quot;&gt;&lt;span style=&quot;--0:#d6deeb;--1:#403f53&quot;&gt;concurrency:&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div class=&quot;ec-line&quot;&gt;&lt;div class=&quot;code&quot;&gt;&lt;span class=&quot;indent&quot;&gt;&lt;span style=&quot;--0:#d6deeb;--1:#403f53&quot;&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;--0:#d6deeb;--1:#403f53&quot;&gt;group: ci-router-${{ github.ref }}&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;div class=&quot;ec-line&quot;&gt;&lt;div class=&quot;code&quot;&gt;&lt;span class=&quot;indent&quot;&gt;&lt;span style=&quot;--0:#d6deeb;--1:#403f53&quot;&gt;  &lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;--0:#d6deeb;--1:#403f53&quot;&gt;cancel-in-progress: true&lt;/span&gt;&lt;/div&gt;&lt;/div&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/figure&gt;&lt;/div&gt;
&lt;p&gt;On &lt;code dir=&quot;auto&quot;&gt;push: branches: [main]&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;ci-router.yml:22-23&lt;/code&gt;) the group key is
&lt;code dir=&quot;auto&quot;&gt;ci-router-refs/heads/main&lt;/code&gt; — &lt;strong&gt;a single group for all of main&lt;/strong&gt; — so each push in
a merge burst cancels the previous main run in-flight. Pushed content lands with
&lt;strong&gt;zero terminal evaluation&lt;/strong&gt;; the first survivor evaluates several merged pushes at
once and misattributes the verdict to whichever head survives (#4347 body; six
cancelled main runs observed 09:07–10:29Z 2026-09-14). PR refs share the same
&lt;code dir=&quot;auto&quot;&gt;cancel-in-progress: true&lt;/code&gt; block but each PR ref is its own key, so PRs
self-coalesce and are unharmed — the damage is specific to the shared
&lt;code dir=&quot;auto&quot;&gt;refs/heads/main&lt;/code&gt; key.&lt;/p&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;#4371 — per-tip &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt; fan-out.&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;.github/workflows/ci-fleet-verdict.yml:3-6&lt;/code&gt;
triggers on &lt;strong&gt;eight upstream workflows&lt;/strong&gt; × &lt;strong&gt;three event types&lt;/strong&gt;
(&lt;code dir=&quot;auto&quot;&gt;types: [requested, in_progress, completed]&lt;/code&gt;), so up to 24 re-fires per upstream
run, and each of those eight workflows itself runs per tip. Measured live: the
&lt;strong&gt;current main tip &lt;code dir=&quot;auto&quot;&gt;e72e13ca01&lt;/code&gt; alone has 60 &lt;code dir=&quot;auto&quot;&gt;CI Fleet Verdict&lt;/code&gt; runs&lt;/strong&gt; in the
last-60 window (LIVE_CI_GROUNDING.md; #4371 quoted 45→61→94→97→82→98). There is &lt;strong&gt;no
top-level &lt;code dir=&quot;auto&quot;&gt;concurrency:&lt;/code&gt; block&lt;/strong&gt; in &lt;code dir=&quot;auto&quot;&gt;ci-fleet-verdict.yml&lt;/code&gt; (verified: file has only
&lt;code dir=&quot;auto&quot;&gt;on:&lt;/code&gt; at &lt;code dir=&quot;auto&quot;&gt;:3&lt;/code&gt; and &lt;code dir=&quot;auto&quot;&gt;jobs:&lt;/code&gt; at &lt;code dir=&quot;auto&quot;&gt;:8&lt;/code&gt;; 94 lines total) and the entry job &lt;code dir=&quot;auto&quot;&gt;identify&lt;/code&gt;
(&lt;code dir=&quot;auto&quot;&gt;:9&lt;/code&gt;) carries none, so every trigger fires a fresh &lt;code dir=&quot;auto&quot;&gt;identify&lt;/code&gt;. Serialization exists
only downstream on &lt;code dir=&quot;auto&quot;&gt;report&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;:42-44&lt;/code&gt;, per-PR group) and &lt;code dir=&quot;auto&quot;&gt;report-main&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;:71-73&lt;/code&gt;,
group &lt;code dir=&quot;auto&quot;&gt;ci-fleet-verdict-main&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;cancel-in-progress: false&lt;/code&gt;). &lt;code dir=&quot;auto&quot;&gt;cancel-in-progress: false&lt;/code&gt; on &lt;code dir=&quot;auto&quot;&gt;report-main&lt;/code&gt; means main-verdict runs &lt;strong&gt;queue rather than coalesce&lt;/strong&gt; —
they pile up faster than they drain (#4371: 96→183 queued in ~5 min) instead of a
survivor dropping the stale ones.&lt;/p&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;#4360-A — main-ledger empty stale-fallback (the COUPLED half of #4360).&lt;/strong&gt; CI
Aggregate triggers on &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt; of “CI Modules” completed
(&lt;code dir=&quot;auto&quot;&gt;.github/workflows/ci-aggregate.yml:55-58&lt;/code&gt;); a &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt; handler always
executes the default-branch file and GitHub labels the run against &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt;. The
fallback source query sets &lt;code dir=&quot;auto&quot;&gt;SOURCE_BRANCH&lt;/code&gt; from the PR head
(&lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml:187&lt;/code&gt;), queries the PR branch for a successful CI Modules run
(&lt;code dir=&quot;auto&quot;&gt;:192-193&lt;/code&gt;), and on empty falls back to &lt;code dir=&quot;auto&quot;&gt;--branch main --status success&lt;/code&gt;
(&lt;code dir=&quot;auto&quot;&gt;:194-199&lt;/code&gt;). When a &lt;strong&gt;red PR-head&lt;/strong&gt; run triggers the main aggregate, the PR query
returns &lt;code dir=&quot;auto&quot;&gt;[]&lt;/code&gt; (PR head was red) and the main query returns &lt;code dir=&quot;auto&quot;&gt;[]&lt;/code&gt; because &lt;strong&gt;main has no
successful CI Modules run of its own&lt;/strong&gt; (on main the CI Modules &lt;code dir=&quot;auto&quot;&gt;report-main&lt;/code&gt; legs are
&lt;code dir=&quot;auto&quot;&gt;skipped&lt;/code&gt;, not &lt;code dir=&quot;auto&quot;&gt;success&lt;/code&gt;). The fail-closed guard then correctly refuses
(&lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml:375-382&lt;/code&gt;, “N registry-expected shard(s) missing from BOTH the
current and fallback runs … refusing to silently treat this run as complete”). &lt;strong&gt;The
guard is right; the fallback source is unsatisfiable&lt;/strong&gt; — live-grounded by “CI
Aggregate = 1 failure” in the recent-25 main window (LIVE_CI_GROUNDING.md).&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Superseded post-Stage-1 — corrected mechanism (2026-09-16).&lt;/strong&gt; The mechanism
narrated in this item is falsified by live verification (Mission
&lt;code dir=&quot;auto&quot;&gt;ci-aggregate-source-eligibility-01M2MFDD&lt;/code&gt;, research D-01/D-02 —
&lt;code dir=&quot;auto&quot;&gt;kitty-specs/ci-aggregate-source-eligibility-01M2MFDD/research.md&lt;/code&gt;). Two corrections,
both verified against the live tree/API:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Mechanism superseded.&lt;/strong&gt; There are &lt;strong&gt;no &lt;code dir=&quot;auto&quot;&gt;report-main&lt;/code&gt; legs&lt;/strong&gt; in &lt;code dir=&quot;auto&quot;&gt;ci-modules.yml&lt;/code&gt;
(&lt;code dir=&quot;auto&quot;&gt;grep report-main&lt;/code&gt; → 0 hits), so the “on main the CI Modules &lt;code dir=&quot;auto&quot;&gt;report-main&lt;/code&gt; legs are
&lt;code dir=&quot;auto&quot;&gt;skipped&lt;/code&gt;, not &lt;code dir=&quot;auto&quot;&gt;success&lt;/code&gt;” narration above is factually wrong. Main CI Modules runs
conclude &lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;success&lt;/code&gt;&lt;/strong&gt; (verified via &lt;code dir=&quot;auto&quot;&gt;gh run list --workflow ci-modules.yml --branch main&lt;/code&gt;), and the &lt;code dir=&quot;auto&quot;&gt;--branch main --status success&lt;/code&gt; fallback query &lt;strong&gt;succeeds&lt;/strong&gt; — it
does &lt;em&gt;not&lt;/em&gt; return &lt;code dir=&quot;auto&quot;&gt;[]&lt;/code&gt;. The refusal is not an unsatisfiable fallback: it is
&lt;code dir=&quot;auto&quot;&gt;reconcile&lt;/code&gt; correctly refusing on &lt;strong&gt;SELECTED-but-undelivered&lt;/strong&gt; shards of a
&lt;strong&gt;red/partial PR-head&lt;/strong&gt; CI Modules run that triggered a &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt;-labelled CI Aggregate.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The failure is cosmetic — no main-verdict consumer reads the mislabelled run.&lt;/strong&gt;
&lt;code dir=&quot;auto&quot;&gt;scripts/ci/fleet_main.py:50&lt;/code&gt; derives the main verdict from an
&lt;code dir=&quot;auto&quot;&gt;event=push&amp;#x26;branch=main&lt;/code&gt; query, to which a &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt;-event aggregate is
invisible; &lt;code dir=&quot;auto&quot;&gt;scripts/ci/fleet_verdict.py:395-405&lt;/code&gt; resolves an aggregate trigger back
to its &lt;strong&gt;source&lt;/strong&gt; CI Modules run and tests provenance on &lt;em&gt;that&lt;/em&gt; run (:405), so a
PR-head aggregate routes to the PR path, never &lt;code dir=&quot;auto&quot;&gt;main=true&lt;/code&gt;; and branch protection
requires only &lt;code dir=&quot;auto&quot;&gt;&quot;Clean install verification&quot;&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;strict:false&lt;/code&gt;) — &lt;strong&gt;CI Aggregate is
not a required check&lt;/strong&gt;. The red run is the PR’s &lt;strong&gt;own correct red&lt;/strong&gt; (its coverage was
incomplete), merely &lt;em&gt;displayed&lt;/em&gt; under a &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; label GitHub assigns to &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt;
handlers and that cannot be relabelled.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;&lt;strong&gt;Consequence.&lt;/strong&gt; #4360-A is therefore a &lt;strong&gt;doctrine-record error plus an untested,
provenance-blind source-selection surface&lt;/strong&gt;, not a release-authority correctness bug.
Axis-3a (below) is delivered as &lt;strong&gt;tested provenance hardening + this doctrine
correction&lt;/strong&gt; (defense-in-depth + honesty), not the removal of a consumed main verdict
false-red. The ratified Decision Outcome and lever table are unchanged; only this
mechanism narration is annotated.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;#4430 — terminal-cancel posts no verdict.&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;scripts/ci/fleet_verdict.py:86-95&lt;/code&gt;:
the red set at &lt;code dir=&quot;auto&quot;&gt;:89&lt;/code&gt; is &lt;code dir=&quot;auto&quot;&gt;{&quot;failure&quot;,&quot;timed_out&quot;,&quot;startup_failure&quot;,&quot;action_required&quot;}&lt;/code&gt;
— &lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;cancelled&lt;/code&gt; is absent&lt;/strong&gt; — and &lt;code dir=&quot;auto&quot;&gt;:95&lt;/code&gt; returns &lt;code dir=&quot;auto&quot;&gt;green&lt;/code&gt; only when all runs are
&lt;code dir=&quot;auto&quot;&gt;completed&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;success&lt;/code&gt;. So a run with &lt;code dir=&quot;auto&quot;&gt;status=completed, conclusion=cancelled&lt;/code&gt; (the
#4399 timeout-kill case) is neither red nor green → &lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;running&lt;/code&gt;&lt;/strong&gt;. The
running-dedup then suppresses the repost: &lt;code dir=&quot;auto&quot;&gt;fleet_verdict.py:325-331&lt;/code&gt; returns without
posting when the new state is &lt;code dir=&quot;auto&quot;&gt;running&lt;/code&gt; and the latest &lt;code dir=&quot;auto&quot;&gt;[ci]&lt;/code&gt; comment already starts
with &lt;code dir=&quot;auto&quot;&gt;[ci] running @&amp;#x3C;sha&gt;&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;:329&lt;/code&gt;). A terminally-cancelled run therefore posts &lt;strong&gt;no
new comment, forever&lt;/strong&gt; — the PR strands on a stale &lt;code dir=&quot;auto&quot;&gt;[ci] running&lt;/code&gt;. The reporter &lt;em&gt;is&lt;/em&gt;
invoked on the cancel (&lt;code dir=&quot;auto&quot;&gt;ci-fleet-verdict.yml:6&lt;/code&gt; includes &lt;code dir=&quot;auto&quot;&gt;completed&lt;/code&gt;); it runs,
computes &lt;code dir=&quot;auto&quot;&gt;running&lt;/code&gt;, and honestly-but-wrongly stays silent because it cannot tell
“still flying” from “terminally cancelled.” (&lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt; / &lt;code dir=&quot;auto&quot;&gt;no suite&lt;/code&gt; exist in the
recognition regex &lt;code dir=&quot;auto&quot;&gt;:323&lt;/code&gt; and vocabulary but are never &lt;em&gt;produced&lt;/em&gt; by &lt;code dir=&quot;auto&quot;&gt;classify&lt;/code&gt;.)&lt;/p&gt;
&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;&lt;strong&gt;Why one decision.&lt;/strong&gt; The candidate menus in the four issue bodies overlap and
interact (Lens B §2, confirmed):&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;#4347 lever “drop cancel-in-progress on main” == #4371 lever (c) “stop cancelling
in-flight main runs.”&lt;/strong&gt; One YAML change satisfies both axes.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;#4371 lever (a) “cap/dedup fleet fan-out per tip” also relieves #4360-A’s ledger
pollution&lt;/strong&gt; — fewer, coalesced main-verdict runs means fewer PR-head runs recorded
against main’s ledger.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;#4371 lever (d) “non-Actions main baseline” would obviate #4360-A entirely&lt;/strong&gt; — a
first-class main-green proof that does not depend on the Actions &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt;
ledger removes the need to fix source-eligibility at all.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;Choosing one lever changes whether the others are still needed. This is a &lt;strong&gt;single
design decision with coupled levers&lt;/strong&gt;, and it is the release-authority contract of ADR
&lt;code dir=&quot;auto&quot;&gt;2026-07-17-1&lt;/code&gt; — which is why the epic marks these children “deliberately NOT promoted
to &lt;code dir=&quot;auto&quot;&gt;status:ready&lt;/code&gt; … the design must settle before implementation” (EPIC_CONTEXT.md:10-11).&lt;/p&gt;
&lt;h3 id=&quot;explicitly-out-of-scope-the-cluster-boundary&quot;&gt;Explicitly out of scope (the cluster boundary)&lt;/h3&gt;
&lt;p&gt;So a reader sees where this ADR stops:&lt;/p&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;#4360-B (diff-scoped shard reconciliation)&lt;/strong&gt; — the operator-flagged live repro (PR
#4448) is a &lt;strong&gt;separable&lt;/strong&gt; defect in the aggregate’s inline reconciler
(&lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml:293-294,336,340-356&lt;/code&gt;) that reproduces on a PR’s &lt;em&gt;own&lt;/em&gt; green head
with no main ledger involved (Lens B §6). It is going into its own mission and is
&lt;strong&gt;not&lt;/strong&gt; decided here.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;#4454&lt;/strong&gt; (tests-only diff selects no module), &lt;strong&gt;#4208&lt;/strong&gt; (router-gate cannot
distinguish timeout-kill from external-cancel), &lt;strong&gt;#4212&lt;/strong&gt; (nightly &lt;code dir=&quot;auto&quot;&gt;set +e&lt;/code&gt; swallow),
&lt;strong&gt;#4374&lt;/strong&gt; (128 out-of-matrix test dirs — human P3 policy) — each a distinct
mechanism, settled or human-owned, shipped separately.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;#4420&lt;/strong&gt; (&lt;code dir=&quot;auto&quot;&gt;status:claimed&lt;/code&gt; + &lt;code dir=&quot;auto&quot;&gt;ready-for-squad&lt;/code&gt; = fleet) and &lt;strong&gt;#4429&lt;/strong&gt; (closed) —
not ours.&lt;/li&gt;
&lt;/ul&gt;
&lt;p&gt;#4208 and #4334 are out of scope but &lt;strong&gt;cross-coupled&lt;/strong&gt; to this cluster; see Consequences.&lt;/p&gt;
&lt;h2 id=&quot;decision-drivers&quot;&gt;Decision Drivers&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;Honesty over convenience (ADR &lt;code dir=&quot;auto&quot;&gt;2026-07-17-1&lt;/code&gt;).&lt;/strong&gt; Green must mean actually proven;
red must mean a real regression. Neither a false-green nor a permanent false-red is
acceptable, and both are live failure modes of the current topology.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Bounded time-to-green-proof.&lt;/strong&gt; A landed main tip must be able to prove itself green
in bounded runner time (#4371’s title verbatim: “Actions lane cannot prove a landed
main tip green in bounded time”).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Single canonical authority (charter).&lt;/strong&gt; One main-green source of truth, not two
competing ledgers.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Preserve the fail-closed guard.&lt;/strong&gt; &lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml:375-382&lt;/code&gt; is correct; only
&lt;em&gt;source-eligibility&lt;/em&gt; may change (#4360: “the fix is in source eligibility, not in
softening the guard”). No option below softens the guard.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Verify where the defect lives.&lt;/strong&gt; These are &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt;-tip behaviours; a fix is only
proven on the merged main tip, never on the PR head that carries it (“a gate never
run is not a gate” — MEMORY).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Human-owned, release-authority-shaping.&lt;/strong&gt; This alters how &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; proves itself
green; the lever selection is the operator’s to ratify.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;considered-options-per-decision-axis&quot;&gt;Considered Options (per decision axis)&lt;/h2&gt;
&lt;p&gt;Four coupled axes. For each, the candidate levers, a &lt;strong&gt;recommended&lt;/strong&gt; lever with
rationale, and the tradeoff of the alternatives.&lt;/p&gt;
&lt;h3 id=&quot;axis-1--main-concurrency-policy-resolves-4347&quot;&gt;Axis 1 — Main-concurrency policy (resolves #4347)&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;1a — Per-SHA group on &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; (RECOMMENDED).&lt;/strong&gt; Change &lt;code dir=&quot;auto&quot;&gt;ci-router.yml:35-37&lt;/code&gt; so the
push-to-main path keys concurrency on &lt;code dir=&quot;auto&quot;&gt;github.sha&lt;/code&gt; (or drops &lt;code dir=&quot;auto&quot;&gt;cancel-in-progress&lt;/code&gt; for
the &lt;code dir=&quot;auto&quot;&gt;refs/heads/main&lt;/code&gt; group), leaving PR refs on the existing per-ref
&lt;code dir=&quot;auto&quot;&gt;cancel-in-progress: true&lt;/code&gt;. Each landed tip then gets its &lt;strong&gt;own&lt;/strong&gt; terminal
evaluation; no push cancels another’s. This is the shared &lt;strong&gt;#4347 lever == #4371(c)&lt;/strong&gt;.
&lt;em&gt;Rationale:&lt;/em&gt; directly removes the cancel-starvation that lands content unevaluated;
minimal, scoped to the main key only; keeps PR self-coalescing intact.
&lt;em&gt;Tradeoff:&lt;/em&gt; more concurrent main runs consume runners — which is exactly why Axis 2
(fan-out cap) must land with it, or the extra main runs multiply the fleet-verdict
storm.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;1b — Merge-gate bookkeeping.&lt;/strong&gt; Record each tip’s verdict in a durable ledger the
merge agent reads, tolerating cancellation. &lt;em&gt;Tradeoff:&lt;/em&gt; larger surface, introduces a
second bookkeeping authority (charter single-authority tension), does not stop the
cancellation itself.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;1c — Merged-tree preflight.&lt;/strong&gt; Evaluate the merged tree before it lands.
&lt;em&gt;Tradeoff:&lt;/em&gt; changes the landing flow, heavier; useful long-term but out of proportion
to the immediate honesty hole.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;1d — Cancelled-run reconcile sweep.&lt;/strong&gt; A sweep re-runs or reconciles cancelled main
runs after the fact. &lt;em&gt;Tradeoff:&lt;/em&gt; reactive, adds a moving part, and leaves the window
where main is unproven.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;axis-2--fleet-verdict-fan-out-cap--dedup-resolves-4371&quot;&gt;Axis 2 — Fleet-verdict fan-out cap / dedup (resolves #4371)&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;2a — Top-level &lt;code dir=&quot;auto&quot;&gt;concurrency&lt;/code&gt; + trim &lt;code dir=&quot;auto&quot;&gt;types:&lt;/code&gt; + dedup (RECOMMENDED).&lt;/strong&gt; Add a
top-level &lt;code dir=&quot;auto&quot;&gt;concurrency:&lt;/code&gt; block to &lt;code dir=&quot;auto&quot;&gt;ci-fleet-verdict.yml&lt;/code&gt; keyed per tip (per-SHA for
main, per-PR for PR heads) so redundant triggers coalesce to a single surviving
&lt;code dir=&quot;auto&quot;&gt;identify&lt;/code&gt;; trim &lt;code dir=&quot;auto&quot;&gt;types: [requested, in_progress, completed]&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;:6&lt;/code&gt;) to &lt;code dir=&quot;auto&quot;&gt;completed&lt;/code&gt;
only — a verdict is a function of &lt;em&gt;terminal&lt;/em&gt; upstream state, so &lt;code dir=&quot;auto&quot;&gt;requested&lt;/code&gt; and
&lt;code dir=&quot;auto&quot;&gt;in_progress&lt;/code&gt; transitions are pure fan-out with no verdict value; and keep/reinforce
the dedup in &lt;code dir=&quot;auto&quot;&gt;scripts/ci/fleet_verdict.py&lt;/code&gt; / &lt;code dir=&quot;auto&quot;&gt;fleet_main.py&lt;/code&gt;. This is &lt;strong&gt;#4371(a)&lt;/strong&gt;.
&lt;em&gt;Rationale:&lt;/em&gt; attacks the multiplicative fan-out at its two roots (event-type triple ×
no throttle); the survivor always re-reads current state, so coalescing loses no
information; &lt;strong&gt;also relieves #4360-A&lt;/strong&gt; by cutting the volume of PR-head runs recorded
against main’s ledger. &lt;em&gt;Tradeoff:&lt;/em&gt; a mis-scoped cap could drop the one verdict a
landed tip needs (a false-red/wedge) — mitigated because the dedup keeps the
&lt;em&gt;survivor&lt;/em&gt; that re-reads current evidence, never drops the last writer; the
&lt;code dir=&quot;auto&quot;&gt;report-main&lt;/code&gt; group must move from &lt;code dir=&quot;auto&quot;&gt;cancel-in-progress: false&lt;/code&gt; (queue) toward
coalesce-with-survivor so the queue cannot outgrow the drain.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;2b — Prioritize main-tip over PR-head runs.&lt;/strong&gt; Give main runs runner priority.
&lt;em&gt;Tradeoff:&lt;/em&gt; does not reduce the fan-out, only reorders it; GitHub-Actions priority
control is coarse.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;2c — Stop cancelling in-flight main runs.&lt;/strong&gt; Identical to Axis 1a — listed by #4371
as its lever (c); adopting 1a satisfies it. Not an independent Axis-2 choice.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;2d — Non-Actions main baseline.&lt;/strong&gt; See Axis 3, lever 3b — a baseline that obviates
the fan-out’s role as the main-green proof.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;axis-3--main-green-source-of-truth-resolves-4360-a&quot;&gt;Axis 3 — Main-green source of truth (resolves #4360-A)&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;3a — Fix source-eligibility, keep the guard (RECOMMENDED near-term).&lt;/strong&gt; Change which
branch/source the fallback query treats as eligible (&lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml:187-199&lt;/code&gt;) so
a PR-head &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt; run labelled against &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; resolves a &lt;strong&gt;main-eligible&lt;/strong&gt;
source or fails with a &lt;em&gt;named&lt;/em&gt; eligibility error — never returns &lt;code dir=&quot;auto&quot;&gt;[]&lt;/code&gt; silently. The
extraction of that inline shell logic (&lt;code dir=&quot;auto&quot;&gt;:191-201&lt;/code&gt;) into a Python helper under
&lt;code dir=&quot;auto&quot;&gt;scripts/ci/&lt;/code&gt; is the &lt;strong&gt;one genuine pytest red-first entry point&lt;/strong&gt; in this cluster.
&lt;strong&gt;The fail-closed guard (&lt;code dir=&quot;auto&quot;&gt;:375-382&lt;/code&gt;) is preserved unchanged.&lt;/strong&gt; &lt;em&gt;Rationale:&lt;/em&gt; smallest
change that makes the main ledger honest; keeps the guard; testable. &lt;em&gt;Tradeoff:&lt;/em&gt; it
keeps main-green &lt;em&gt;dependent on the Actions ledger&lt;/em&gt; — the deeper fragility #4371(d)
removes. Contingency: if Axis 2 (fan-out cap) is chosen such that main produces a
first-class successful CI Modules run of its own, the fallback becomes satisfiable and
3a’s role shrinks to a safety net.&lt;/p&gt;
&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Superseded post-Stage-1 — corrected mechanism (2026-09-16).&lt;/strong&gt; The near-term premise
for 3a is corrected by live verification (Mission
&lt;code dir=&quot;auto&quot;&gt;ci-aggregate-source-eligibility-01M2MFDD&lt;/code&gt;, research D-01/D-02 —
&lt;code dir=&quot;auto&quot;&gt;kitty-specs/ci-aggregate-source-eligibility-01M2MFDD/research.md&lt;/code&gt;), matching the
annotation on #4360-A in &lt;em&gt;Context and Problem Statement&lt;/em&gt; above. Two corrections:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;Mechanism superseded.&lt;/strong&gt; The fallback does &lt;strong&gt;not&lt;/strong&gt; return &lt;code dir=&quot;auto&quot;&gt;[]&lt;/code&gt;: there are &lt;strong&gt;no
&lt;code dir=&quot;auto&quot;&gt;report-main&lt;/code&gt; legs&lt;/strong&gt; in &lt;code dir=&quot;auto&quot;&gt;ci-modules.yml&lt;/code&gt; (0 grep hits), main CI Modules runs conclude
&lt;strong&gt;&lt;code dir=&quot;auto&quot;&gt;success&lt;/code&gt;&lt;/strong&gt;, and the &lt;code dir=&quot;auto&quot;&gt;--branch main --status success&lt;/code&gt; query &lt;strong&gt;succeeds&lt;/strong&gt;. The
refusal is &lt;code dir=&quot;auto&quot;&gt;reconcile&lt;/code&gt; on &lt;strong&gt;SELECTED-but-undelivered&lt;/strong&gt; shards of a &lt;strong&gt;red/partial
PR-head&lt;/strong&gt; run adjudicated under a &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; label — not an unsatisfiable source query.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;The failure is cosmetic.&lt;/strong&gt; No main-verdict consumer reads the mislabelled run
(&lt;code dir=&quot;auto&quot;&gt;fleet_main.py:50&lt;/code&gt; &lt;code dir=&quot;auto&quot;&gt;event=push&amp;#x26;branch=main&lt;/code&gt;; &lt;code dir=&quot;auto&quot;&gt;fleet_verdict.py:395-405&lt;/code&gt; resolves the
aggregate back to its source run and tests provenance there, never &lt;code dir=&quot;auto&quot;&gt;main=true&lt;/code&gt;; branch
protection requires only &lt;code dir=&quot;auto&quot;&gt;&quot;Clean install verification&quot;&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;strict:false&lt;/code&gt; — CI Aggregate
is not a required check). It is the PR’s own correct red under a &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; label.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;&lt;strong&gt;Therefore Axis-3a is delivered as tested provenance hardening + this doctrine
correction&lt;/strong&gt; — defense-in-depth + honesty: PR-head/failed evidence is provably never
treated as an eligible &lt;em&gt;main&lt;/em&gt; source, and the source decision always resolves either an
eligible source or a &lt;strong&gt;named&lt;/strong&gt; reason instead of a silent &lt;code dir=&quot;auto&quot;&gt;[]&lt;/code&gt; — &lt;strong&gt;not&lt;/strong&gt; a
release-authority correctness fix (no consumed main-verdict false-red exists to remove).
The ratified lever selection (3a near-term behind the preserved guard; 3b as the durable
target) is &lt;strong&gt;unchanged&lt;/strong&gt;, and the fail-closed guard (&lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml:375-382&lt;/code&gt;) stays
untouched.&lt;/p&gt;
&lt;/blockquote&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;3b — Stand up a non-Actions main baseline (RECOMMENDED as the durable target).&lt;/strong&gt;
An authoritative main-green proof that does not depend on the &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt; ledger
(&lt;strong&gt;#4371(d)&lt;/strong&gt;) — this &lt;strong&gt;obviates #4360-A entirely&lt;/strong&gt;. &lt;em&gt;Rationale:&lt;/em&gt; removes the “PR-head
runs are inputs to a main gate” shape at the root. &lt;em&gt;Tradeoff:&lt;/em&gt; a larger build (a new
producer/store and its own trust model); higher risk if rushed. &lt;strong&gt;Recommendation:
adopt 3a now behind the preserved guard, and record 3b as the target that retires
this axis&lt;/strong&gt; — do not soften the guard to bridge the gap.&lt;/p&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;p&gt;&lt;strong&gt;3c — Exclude PR-branch &lt;code dir=&quot;auto&quot;&gt;workflow_run&lt;/code&gt; from the main ledger.&lt;/strong&gt; Filter the aggregate
so PR-head runs never post against main. &lt;em&gt;Tradeoff:&lt;/em&gt; correct in spirit but leaves
main with &lt;em&gt;no&lt;/em&gt; proof source until 3a/3b supplies one — risks a permanent false-red on
main. Best combined with 3a, not alone.&lt;/p&gt;
&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;axis-4--terminal-cancel-verdict-class--consumer-resolves-4430&quot;&gt;Axis 4 — Terminal-cancel verdict class + consumer (resolves #4430)&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;4a — In-CI &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt; terminal verdict, releases the head (RECOMMENDED).&lt;/strong&gt; Add a
terminal-cancel class to &lt;code dir=&quot;auto&quot;&gt;classify&lt;/code&gt; (&lt;code dir=&quot;auto&quot;&gt;fleet_verdict.py:86-95&lt;/code&gt;): when a required run is
&lt;code dir=&quot;auto&quot;&gt;completed&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;cancelled&lt;/code&gt; (and not already red via &lt;code dir=&quot;auto&quot;&gt;timed_out&lt;/code&gt;), emit &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt; —
distinct from both &lt;code dir=&quot;auto&quot;&gt;running&lt;/code&gt; and &lt;code dir=&quot;auto&quot;&gt;green&lt;/code&gt; — and ensure the dedup (&lt;code dir=&quot;auto&quot;&gt;:325-331&lt;/code&gt;) does
&lt;strong&gt;not&lt;/strong&gt; suppress it (the running-guard at &lt;code dir=&quot;auto&quot;&gt;:329&lt;/code&gt; matches only &lt;code dir=&quot;auto&quot;&gt;state==&quot;running&quot;&lt;/code&gt;, so a
fresh &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt; comment posts and releases the head as re-triggerable). Thread the
same class through the main path (&lt;code dir=&quot;auto&quot;&gt;scripts/ci/fleet_main.py&lt;/code&gt;,
&lt;code dir=&quot;auto&quot;&gt;ci-fleet-verdict.yml:92-93&lt;/code&gt;) for one coherent “terminal-cancel = infra-error, never
green, always released” semantics across PR and main. &lt;em&gt;Rationale:&lt;/em&gt; smallest honest
fix; preserves the never-green safety (&lt;code dir=&quot;auto&quot;&gt;fleet_verdict.py:87&lt;/code&gt;); un-strands the PR.
RED-FIRST via &lt;code dir=&quot;auto&quot;&gt;tests/ci/test_fleet_verdict.py&lt;/code&gt; (flip the pinned
&lt;code dir=&quot;auto&quot;&gt;(&quot;completed&quot;,&quot;cancelled&quot;,&quot;running&quot;)&lt;/code&gt; contract at &lt;code dir=&quot;auto&quot;&gt;:155&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;:289-290&lt;/code&gt; to a non-green,
non-running class, plus a new &lt;code dir=&quot;auto&quot;&gt;test_terminal_cancel_reposts_and_is_not_suppressed&lt;/code&gt;).
&lt;em&gt;Tradeoff:&lt;/em&gt; &lt;code dir=&quot;auto&quot;&gt;classify&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;report&lt;/code&gt; are the release-authority gate for every PR; a wrong
change here could green-wash or spuriously red — mitigated by keeping &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt;
strictly outside the green path.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;4b — Merge-agent sweep surfaces the stale-running as a watch item.&lt;/strong&gt; A consumer
(merge/night-watch agent) flags “last verdict is &lt;code dir=&quot;auto&quot;&gt;running&lt;/code&gt; but its run is terminal.”
&lt;em&gt;Tradeoff:&lt;/em&gt; lives outside CI (&lt;code dir=&quot;auto&quot;&gt;agents/merge.md&lt;/code&gt; surface), reactive, does not release
the head automatically. &lt;em&gt;Recommendation:&lt;/em&gt; pair as a backstop to 4a, not as the
primary — the head should be released in-CI, not by a human/agent sweep.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;decision-outcome-ratified-2026-09-15&quot;&gt;Decision Outcome (ratified 2026-09-15)&lt;/h2&gt;
&lt;p&gt;&lt;strong&gt;Ratified coherent bundle&lt;/strong&gt; (the levers are chosen &lt;em&gt;together&lt;/em&gt; precisely because they
interact; the operator ratified all four recommended levers on 2026-09-15):&lt;/p&gt;



































&lt;table&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;Axis&lt;/th&gt;&lt;th&gt;Recommended lever&lt;/th&gt;&lt;th&gt;Resolves&lt;/th&gt;&lt;th&gt;Coupling note&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;1 — main-concurrency&lt;/td&gt;&lt;td&gt;&lt;strong&gt;1a&lt;/strong&gt; per-SHA group on &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt;&lt;/td&gt;&lt;td&gt;#4347&lt;/td&gt;&lt;td&gt;== #4371(c)&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;2 — fleet fan-out&lt;/td&gt;&lt;td&gt;&lt;strong&gt;2a&lt;/strong&gt; top-level &lt;code dir=&quot;auto&quot;&gt;concurrency&lt;/code&gt; + trim &lt;code dir=&quot;auto&quot;&gt;types:&lt;/code&gt; to &lt;code dir=&quot;auto&quot;&gt;completed&lt;/code&gt; + dedup&lt;/td&gt;&lt;td&gt;#4371&lt;/td&gt;&lt;td&gt;relieves #4360-A&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;3 — main-green source&lt;/td&gt;&lt;td&gt;&lt;strong&gt;3a&lt;/strong&gt; fix source-eligibility, &lt;strong&gt;guard preserved&lt;/strong&gt;; &lt;strong&gt;3b&lt;/strong&gt; non-Actions baseline as the target that retires the axis&lt;/td&gt;&lt;td&gt;#4360-A&lt;/td&gt;&lt;td&gt;#4371(d) obviates 3a&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;4 — terminal-cancel&lt;/td&gt;&lt;td&gt;&lt;strong&gt;4a&lt;/strong&gt; in-CI &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt;, releases head; 4b sweep as backstop&lt;/td&gt;&lt;td&gt;#4430&lt;/td&gt;&lt;td&gt;shares fleet-verdict surface with Axis 2&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;
&lt;p&gt;&lt;strong&gt;Why this bundle and not four point fixes.&lt;/strong&gt; 1a removes cancel-starvation but &lt;em&gt;adds&lt;/em&gt;
main runs; 2a caps the fan-out those extra runs would otherwise multiply — so 1a and 2a
must land together. 2a’s coalescing cuts PR-head pollution of main’s ledger, shrinking
what 3a must repair. 3b, if built, retires Axis 3 outright. 4a shares the exact
&lt;code dir=&quot;auto&quot;&gt;classify&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;report&lt;/code&gt; surface 2a’s dedup touches, so they must be one coherent
terminal-state semantics. A point fix on any one silently reopens another (#4371).&lt;/p&gt;
&lt;h3 id=&quot;the-two-bounding-failure-modes-and-how-the-bundle-avoids-each&quot;&gt;The two bounding failure modes, and how the bundle avoids each&lt;/h3&gt;
&lt;p&gt;Every option is bounded by two outcomes, both worse than the status quo (Lens B §4):&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;&lt;strong&gt;False-green — the fleet merges unproven code.&lt;/strong&gt; Avoided by: &lt;strong&gt;keeping the
fail-closed guard&lt;/strong&gt; (&lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml:375-382&lt;/code&gt;) untouched — only source-eligibility
(3a) changes; &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt; (4a) is defined strictly outside the green path
(&lt;code dir=&quot;auto&quot;&gt;fleet_verdict.py:87&lt;/code&gt; never-green invariant preserved); no PR-head partial-shard set
is ever allowed to satisfy the main ledger. This upholds ADR &lt;code dir=&quot;auto&quot;&gt;2026-07-17-1&lt;/code&gt;: green
still means actually proven.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;False-red / permanent-wedge — main never proves green.&lt;/strong&gt; Avoided by: 1a giving
each landed tip its own terminal evaluation (no cancel-starvation); 2a’s dedup
keeping the &lt;em&gt;survivor&lt;/em&gt; that re-reads current evidence (never dropping the last-writer
verdict a landed tip needs); 4a posting &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt; to &lt;em&gt;release&lt;/em&gt; a
terminally-cancelled head instead of stranding it on &lt;code dir=&quot;auto&quot;&gt;[ci] running&lt;/code&gt;; and 3a resolving
a named eligibility error instead of an empty &lt;code dir=&quot;auto&quot;&gt;[]&lt;/code&gt;.&lt;/li&gt;
&lt;/ol&gt;
&lt;p&gt;The recommendation deliberately refuses the two tempting shortcuts that would break (1):
softening the aggregate guard, and letting a PR-head run’s partial shards count as
main-green. Both are explicitly rejected.&lt;/p&gt;
&lt;h2 id=&quot;consequences&quot;&gt;Consequences&lt;/h2&gt;
&lt;h3 id=&quot;positive&quot;&gt;Positive&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;A landed main tip can prove itself green in bounded runner time; the 60-runs-per-tip
fan-out (LIVE_CI_GROUNDING.md) collapses toward one verdict per tip.&lt;/li&gt;
&lt;li&gt;Merge bursts no longer land content on main unevaluated (#4347 closed).&lt;/li&gt;
&lt;li&gt;The main ledger stops being fed by PR-head runs, or fails with a &lt;em&gt;named&lt;/em&gt; error instead
of a silent &lt;code dir=&quot;auto&quot;&gt;[]&lt;/code&gt; (#4360-A closed near-term; #4371(d) path recorded for durable close).&lt;/li&gt;
&lt;li&gt;A terminally-cancelled PR is released and re-triggerable instead of wedged on &lt;code dir=&quot;auto&quot;&gt;[ci] running&lt;/code&gt; (#4430 closed).&lt;/li&gt;
&lt;li&gt;Single main-green authority preserved; the fail-closed guard is untouched, so
release-authority honesty (ADR &lt;code dir=&quot;auto&quot;&gt;2026-07-17-1&lt;/code&gt;) is strengthened, not weakened.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;negative--accepted-risk&quot;&gt;Negative / accepted risk&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;The concurrency (1a) and fan-out (2a) changes are &lt;strong&gt;config-level YAML&lt;/strong&gt;, largely not
unit-testable in pytest; provable only by workflow-lint / golden-YAML shape assertions
plus dedup unit tests in &lt;code dir=&quot;auto&quot;&gt;tests/ci/&lt;/code&gt;. Only Axis 3a’s extracted eligibility helper has
a genuine pytest red-first pin. The mission spec must state this honestly rather than
pretend a unit test covers the YAML change (Lens B §3).&lt;/li&gt;
&lt;li&gt;&lt;code dir=&quot;auto&quot;&gt;classify&lt;/code&gt;/&lt;code dir=&quot;auto&quot;&gt;report&lt;/code&gt; (Axis 4a) is the per-PR release gate; the change adds a new
non-green class and must not widen the green path. MEDIUM risk; the never-green tests
(&lt;code dir=&quot;auto&quot;&gt;fleet_verdict.py:87&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;test_truncated_files_and_deferred_pr_never_green&lt;/code&gt;) must stay
green.&lt;/li&gt;
&lt;li&gt;1a increases concurrent main runs; without 2a landing with it, the fleet-verdict storm
worsens. The two are sequenced together for this reason.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;staged-rollout--a-gate-never-run-is-not-a-gate&quot;&gt;Staged rollout — “a gate never run is not a gate”&lt;/h3&gt;
&lt;p&gt;These are &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt;-tip behaviours. The change alters the very lane that would validate the
PR carrying it, so it &lt;strong&gt;can only be fully proven on &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt; after merge&lt;/strong&gt; — the first
moment the main-tip path exists (Lens B §4; MEMORY “a gate never run is not a gate”).
The implementing mission must therefore:&lt;/p&gt;
&lt;ol&gt;
&lt;li&gt;Land the levers in a &lt;strong&gt;staged sequence&lt;/strong&gt; (concurrency + fan-out cap together, then
source-eligibility, then terminal-cancel), each on an &lt;code dir=&quot;auto&quot;&gt;issue-&amp;#x3C;n&gt;-&amp;#x3C;slug&gt;&lt;/code&gt; branch → PR
→ operator merges (never a push to &lt;code dir=&quot;auto&quot;&gt;main&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;Verify each stage on the merged main tip&lt;/strong&gt;, not on the PR head — count fleet-verdict
runs per tip, confirm a single coalesced verdict, confirm the aggregate resolves a
main-eligible source, confirm a forced cancel posts &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Treat any consumer this ADR did not name as a &lt;strong&gt;finding against this ADR&lt;/strong&gt;, not a
licence to improvise around it.&lt;/li&gt;
&lt;/ol&gt;
&lt;h3 id=&quot;cross-coupling-out-of-scope-issues-this-cluster-moves&quot;&gt;Cross-coupling (out-of-scope issues this cluster moves)&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;&lt;strong&gt;#4208 (its inputs).&lt;/strong&gt; #4347’s &lt;code dir=&quot;auto&quot;&gt;cancel-in-progress&lt;/code&gt; is what &lt;em&gt;produces&lt;/em&gt; the
&lt;code dir=&quot;auto&quot;&gt;cancelled&lt;/code&gt; states #4208’s router-gate must classify. Adopting Axis 1a &lt;strong&gt;stops
cancelling main runs, changing #4208’s input distribution&lt;/strong&gt; — the two must be
coordinated so #4208’s timeout-vs-external-cancel classifier is not built against a
distribution this ADR removes.&lt;/li&gt;
&lt;li&gt;&lt;strong&gt;#4334 (sonar-pr consumes aggregate shards).&lt;/strong&gt; The &lt;code dir=&quot;auto&quot;&gt;sonar-pr&lt;/code&gt; job in
&lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml&lt;/code&gt; consumes the reconciled shard coverage; it is &lt;code dir=&quot;auto&quot;&gt;continue-on-error&lt;/code&gt;
and excluded from the terminal &lt;code dir=&quot;auto&quot;&gt;aggregate-gate&lt;/code&gt; (reported, not required). An Axis-3a
source-eligibility change must be verified to leave &lt;code dir=&quot;auto&quot;&gt;sonar-pr&lt;/code&gt; tolerant of the
resolved source set — non-blocking, so a smaller/different set is not a false-red on
the PR path, but confirm it does not error the informational job. &lt;strong&gt;[HYP]&lt;/strong&gt; — sonar-pr
not re-read this pass.&lt;/li&gt;
&lt;/ul&gt;
&lt;h3 id=&quot;neutral&quot;&gt;Neutral&lt;/h3&gt;
&lt;ul&gt;
&lt;li&gt;No doctrine is changed. ADR &lt;code dir=&quot;auto&quot;&gt;2026-07-17-1&lt;/code&gt; remains in force and unedited; this ADR
operationalises its “green means proven” invariant on the main-tip topology.&lt;/li&gt;
&lt;li&gt;The diff-scoped reconciler (#4360-B) is untouched and proceeds as its own mission;
this ADR’s Axis-3 changes do not read the reconciler’s selection logic.&lt;/li&gt;
&lt;/ul&gt;
&lt;h2 id=&quot;confirmation&quot;&gt;Confirmation&lt;/h2&gt;
&lt;p&gt;Ratification of this ADR is confirmed in practice when, on the merged main tip:
(a) one landed tip produces a bounded, single coalesced &lt;code dir=&quot;auto&quot;&gt;CI Fleet Verdict&lt;/code&gt; rather than
dozens; (b) no main content lands without a terminal evaluation during a merge burst;
(c) CI Aggregate on main resolves a main-eligible source or fails with a named
eligibility error, with the fail-closed guard (&lt;code dir=&quot;auto&quot;&gt;ci-aggregate.yml:375-382&lt;/code&gt;) unchanged;
and (d) a terminally-cancelled PR-head run posts &lt;code dir=&quot;auto&quot;&gt;infra-error&lt;/code&gt; and releases the head.
A main tip reported green on unproven content, or a permanently-wedged main verdict,
is a violation of this ADR and of ADR &lt;code dir=&quot;auto&quot;&gt;2026-07-17-1&lt;/code&gt;.&lt;/p&gt;
&lt;p&gt;Confidence: &lt;strong&gt;high&lt;/strong&gt; on the mechanisms (every &lt;code dir=&quot;auto&quot;&gt;file:line&lt;/code&gt; re-verified in-tree at HEAD
&lt;code dir=&quot;auto&quot;&gt;e72e13ca01&lt;/code&gt;); &lt;strong&gt;medium&lt;/strong&gt; on the exact YAML/dedup mechanics, which are the implementing
mission’s to finalise within the lever selection ratified here.&lt;/p&gt;
&lt;h2 id=&quot;references&quot;&gt;References&lt;/h2&gt;
&lt;ul&gt;
&lt;li&gt;Issues: &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4437&quot;&gt;#4437&lt;/a&gt; (epic),
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4347&quot;&gt;#4347&lt;/a&gt;,
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4371&quot;&gt;#4371&lt;/a&gt;,
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4360&quot;&gt;#4360&lt;/a&gt; (part A — the coupled
main-ledger half; part B out of scope),
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4430&quot;&gt;#4430&lt;/a&gt;. Cross-coupled but out
of scope: &lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4208&quot;&gt;#4208&lt;/a&gt;,
&lt;a href=&quot;https://github.com/spec-kitty/spec-kitty/issues/4334&quot;&gt;#4334&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;Governing ADR: &lt;a href=&quot;/adr/3.x/2026-07-17-1-red-main-is-honest-ci-is-release-authority/&quot;&gt;&lt;code dir=&quot;auto&quot;&gt;2026-07-17-1-red-main-is-honest-ci-is-release-authority.md&lt;/code&gt;&lt;/a&gt;.&lt;/li&gt;
&lt;li&gt;Research inputs (local, &lt;code dir=&quot;auto&quot;&gt;work/&lt;/code&gt;): &lt;code dir=&quot;auto&quot;&gt;EPIC_CONTEXT.md&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;SYNTHESIS.md&lt;/code&gt;,
&lt;code dir=&quot;auto&quot;&gt;LIVE_CI_GROUNDING.md&lt;/code&gt;, &lt;code dir=&quot;auto&quot;&gt;lenses/B-fanout-cluster.md&lt;/code&gt; (primary),
&lt;code dir=&quot;auto&quot;&gt;lenses/C-verdict-and-nightly.md&lt;/code&gt; (#4430).&lt;/li&gt;
&lt;li&gt;Code anchors (HEAD &lt;code dir=&quot;auto&quot;&gt;e72e13ca01&lt;/code&gt;): &lt;code dir=&quot;auto&quot;&gt;.github/workflows/ci-router.yml:22-23,35-37&lt;/code&gt;;
&lt;code dir=&quot;auto&quot;&gt;.github/workflows/ci-fleet-verdict.yml:3-6,42-44,71-73,92-93&lt;/code&gt;;
&lt;code dir=&quot;auto&quot;&gt;.github/workflows/ci-aggregate.yml:55-58,187-199,375-382&lt;/code&gt;;
&lt;code dir=&quot;auto&quot;&gt;scripts/ci/fleet_verdict.py:86-95,323-331&lt;/code&gt;; &lt;code dir=&quot;auto&quot;&gt;scripts/ci/fleet_main.py&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;</content>
</entry>
