fix(OMN-16808): the autoclose sweep must not repeat itself on a ticket it already commented on

.github/workflows/evidence-autoclose-sweep.yml

@@ -67,7 +67,40 @@ on:
       lookback_hours:
         description: 'Scan companions merged within this many hours of now.'
         required: false
-        default: '2'
+        # 6h, raised from 2h (OMN-16792, landed with OMN-16808).
+        #
+        # `schedule` is best-effort: GitHub drops ticks under load rather than
+        # queueing them. Measured over 2026-08-22T20:03Z -> 2026-08-27T19:05Z,
+        # 194 scheduled runs fired against 238 expected at `*/30` (81%), and
+        # the discriminating measurement is createdAt->startedAt = 0s median
+        # AND 0s max across all 194: nothing was slow, the missing runs were
+        # never created. Observed inter-run gaps: median 30 min, p90 43 min,
+        # p100 236 min (3h56m).
+        #
+        # A 2h lookback is smaller than that p100 gap, so a companion merged
+        # inside the uncovered window was scanned by NO run, ever — the sweep
+        # only looks backwards a fixed window and never revisits. 6h clears the
+        # observed p100 by ~1.5x, so a dropped tick is absorbed rather than
+        # lost.
+        #
+        # The reason this is a safe one-line change NOW and was not before:
+        # widening the window widens the OVERLAP between consecutive runs, and
+        # before OMN-16808 every overlapping run re-posted an identical gap
+        # comment on every gapped ticket it re-saw. The read-before-write dedup
+        # in the handler makes re-scanning a companion cost nothing observable,
+        # so the lookback can be sized for the worst measured gap instead of
+        # for comment noise. That is OMN-16792 option 1, chosen because it
+        # removes the data-loss risk without deciding anything architectural;
+        # options 2 (event trigger) and 3 (runtime scheduler) stay open on that
+        # ticket.
+        #
+        # Residual, recorded not fixed here: `max_companions` is capped at 50
+        # per run. A 10h window hit that cap on 2026-08-28 (run 33128661860,
+        # companions_scanned=50), so a window wide enough to truncate would
+        # reintroduce the same loss by a different door. At the observed OCC
+        # merge rate (~50 per 10h) a 6h window lands around 30 companions,
+        # under the cap. Sizing the cap belongs to OMN-16792.
+        default: '6'
       diagnose_tickets:
         description: >-
           OMN-16788 diagnostic. Comma-separated ticket ids (e.g.
@@ -481,7 +514,7 @@ jobs:
             exit 1
           fi
           uv run onex skill evidence_autoclose_sweep \
-            --lookback-hours "${{ github.event.inputs.lookback_hours || '2' }}"
+            --lookback-hours "${{ github.event.inputs.lookback_hours || '6' }}"
 
       - name: Post summary
         if: always()

docker/runners/runner-image.lock.json

@@ -1,12 +1,12 @@
 {
   "base_image_digest": "sha256:3ba65aa20f86a0fad9df2b2c259c613df006b2e6d0bfcc8a146afb8c525a9751",
   "gh_version": "2.67.0",
-  "identity_digest": "47224bdefe93869e4dd05501c506eb18",
+  "identity_digest": "7fae4af5f1fb6695a6e032a366e12613",
   "image_version": 7,
   "kubectl_version": "1.32.1",
   "python_version": "3.12",
   "runner_version": "2.336.0",
-  "shared_env_digest": "ac02631ce31562210f83ee2f",
+  "shared_env_digest": "a7a173a77e4d15465404b3b3",
   "shared_env_install_args": "--frozen --all-extras --all-groups --no-install-project",
   "uv_version": "0.6.14"
 }

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "omnibase_infra"
-version = "0.38.12"
+version = "0.38.13"
 description = "ONEX Infrastructure - Service integration and database infrastructure tools"
 authors = [{ name = "OmniNode.ai", email = "contact@omninode.ai" }]
 license = {text = "MIT"}

src/omnibase_infra/nodes/node_evidence_autoclose_sweep_effect/contract.yaml

@@ -139,6 +139,18 @@ error_handling:
       description: >
         The verdict carried no `result.terminal_payload.behavior_proving_count` — the verifier predates OMN-15911 and cannot say whether any green check proved behavior. Fails closed as ERROR_VERIFY_UNPARSEABLE rather than inferring a value: reading the absence as "zero behavior proof" would post a Linear comment stating a finding the run never made, and reading it as satisfied would flip on exactly the arithmetic the guard exists to stop.
 
+      recoverable: true
+      retry_strategy: "none"
+    - name: "DUPLICATE_SWEEP_COMMENT"
+      description: >
+        The sweep reached a gap verdict it has ALREADY posted on this ticket and did not repeat it (OMN-16808). Enumeration is a bare `now - lookback_hours` window with no cursor, so one merged companion sits inside several consecutive scheduled windows; before this guard every one of them posted an identical comment. Every sweep comment now carries a stable marker keyed on (gap class, verdict fingerprint) — NOT on the companion PR, because the same verdict re-derived from a later companion is the same statement — and the ticket's existing comments are read before any write. Recorded as SKIPPED_DUPLICATE_COMMENT and counted under tickets_skipped: the gap is real and still open, but this run added no information, and counting it under tickets_gap_posted would report comments that were never written. A CHANGED verdict has a different fingerprint and IS posted.
+
+      recoverable: false
+      retry_strategy: "none"
+    - name: "COMMENT_HISTORY_UNREADABLE"
+      description: >
+        The ticket's existing comments could not be read (Linear API failure, an uninterpretable payload, or a history longer than the paginated read cap), so the sweep cannot establish whether it has already made this statement. Fails closed as ERROR_LINEAR_API and writes nothing (OMN-16808): "I could not check" must never resolve to "so I will post". The flip is withheld on this path regardless — it is only ever reached from a gap classification.
+
       recoverable: true
       retry_strategy: "none"
     - name: "GITHUB_FILES_PAYLOAD_UNREADABLE"