diff --git a/CHANGELOG.md b/CHANGELOG.md
index bf192f6a9e..52f7f3c807 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,3 +1,13 @@
+### Agent-mention sweep regression follows completion order
+
+- Replace the repository-order assertion with a deterministic completion-order
+  fixture for `list_recent_pull_requests`. The production iterator already
+  uses `concurrent.futures.as_completed`; the stale test could pass or fail
+  with thread scheduling and contradicted the latency boundary that a completed
+  repository must be yielded without waiting for a slower sibling. The fixture
+  blocks the first repository until the caller has received the second, while
+  retaining the two-worker ceiling.
+
 ### Failed-check finding names the Strix sandbox instead of the gateway
 
 - `opencode-review-dispatch.yml`'s `emit_strix_provider_failure_finding` rendered one fixed finding for every `STRIX_PROVIDER_UNAVAILABLE` line, whose Root cause read "The contextual-orchestrator gateway or its discovered provider pool was unavailable for this run". `#1953` had just given the Strix sandbox bootstrap failure its own second verdict token (`STRIX_SANDBOX_UNAVAILABLE`) precisely because that attribution is wrong for it -- the sandbox container never reaches its Caido proxy, so the run dies before the gateway serves anything -- and this consumer re-applied the wrong attribution one step downstream, into the review findings and the failure census. The emitter now branches on the second token: a sandbox verdict gets a finding that names Strix's sandbox, says the verdict does not name the gateway, and tells the reader not to change gateway or provider configuration on its strength. A `STRIX_PROVIDER_UNAVAILABLE` line without the token keeps its existing text verbatim, so the gateway class has no regression surface. No test covered this finding text at all before (`gateway or its discovered provider pool` matched nothing under `tests/`); `tests/test_opencode_dispatch_strix_sandbox_finding.py` now runs the production emitter from the published run block and pins both directions plus the no-signal case. Refs #1953, #1935.
diff --git a/docs/doctoring/agent-mention-sweep-completion-order.md b/docs/doctoring/agent-mention-sweep-completion-order.md
new file mode 100644
index 0000000000..62585fe2ec
--- /dev/null
+++ b/docs/doctoring/agent-mention-sweep-completion-order.md
@@ -0,0 +1,46 @@
+# Agent-mention sweep completion-order regression
+
+Decision date: **2026-09-08**
+
+## Problem
+
+`list_recent_pull_requests` submits bounded repository fetches concurrently and
+iterates them with `concurrent.futures.as_completed`. Its regression instead
+claimed that results remain in repository order and asserted
+`first, second` without controlling completion. The assertion therefore tested
+thread timing rather than the production contract.
+
+A slow first repository is an operational failure scene: if the already-finished
+second repository is hidden behind it, the sweep wastes its bounded dispatch
+window and may omit actionable mentions before the next rotation.
+
+## Decision
+
+Keep the production implementation unchanged. Replace only the stale regression
+with a deterministic fixture. The first repository waits on an event; the caller
+must receive the second repository before releasing that event. The test retains
+the exact two-worker assertion.
+
+This is the minimum owner repair. It adds no scheduler abstraction, timeout,
+retry, ordering buffer, or dependency.
+
+## Evidence
+
+- RED `570da463b0ce7f4837727a91557187624a7b37db` makes the completion order
+  deterministic while retaining the obsolete `first, second` expectation.
+  The production iterator necessarily yields `second` first.
+- GREEN replaces the expectation with `second, first` and names the latency
+  contract directly.
+- Exact-head hosted Checks remain the admission authority.
+
+## Risks and rollback
+
+The event wait has a 30-second failure ceiling so a broken executor produces an
+actionable failure instead of hanging the suite. Rollback is removal of this
+test-only successor; production behavior is unchanged.
+
+## References
+
+Python Software Foundation. (2026). *concurrent.futures — Launching parallel
+tasks*. Python 3 documentation.
+https://docs.python.org/3/library/concurrent.futures.html#concurrent.futures.as_completed
diff --git a/docs/product-technical-gap-baseline.md b/docs/product-technical-gap-baseline.md
index 1cc9e20313..94d2759685 100644
--- a/docs/product-technical-gap-baseline.md
+++ b/docs/product-technical-gap-baseline.md
@@ -1,5 +1,20 @@
 # Product and Technical Gap Baseline
 
+## Exact-head amendment — 2026-09-08
+
+- **Gap:** the agent-mention sweep production iterator yields repository results
+  with `concurrent.futures.as_completed`, but its current-main regression
+  claimed and asserted repository order. That scheduler-dependent assertion
+  could conceal the intended latency boundary or fail nondeterministically.
+- **Action/status:** Proposed on
+  `fix/agent-sweep-completion-order-regression`. A deterministic two-repository
+  fixture holds the first request until the already-completed second result is
+  observed, then asserts second-before-first and preserves the two-worker
+  ceiling. Protected-main integration and exact-head Checks remain required.
+- **Boundary:** no production scheduling, retry, credential, merge, or scanner
+  policy changes.
+
+
 작성 기준일: **2026-08-26 10:35 KST**
 대상: **ContextualWisdomLab/.github** 중앙 거버넌스·자동화 레포지터리와 이를 소비하는 naruon 생태계
 현재 보호된 `main`: `826b92394c63deb6981c3a8d16a724d71f85a0d7`
diff --git a/tests/test_agent_mention_sweep_regressions.py b/tests/test_agent_mention_sweep_regressions.py
index 643f562cc0..217bfcebc0 100644
--- a/tests/test_agent_mention_sweep_regressions.py
+++ b/tests/test_agent_mention_sweep_regressions.py
@@ -4,6 +4,7 @@
 
 import importlib
 import sys
+import threading
 from datetime import datetime, timezone
 from pathlib import Path
 
@@ -94,11 +95,25 @@ def test_pull_pagination_stops_at_cutoff_without_loading_later_pages() -> None:
     assert sweep.flatten_pages([{"number": 1}]) == [{"number": 1}]
 
 
-def test_recent_pull_requests_use_bounded_parallel_repository_fetches(monkeypatch) -> None:
-    """Repository fetches are parallel but results remain repository ordered."""
+def test_recent_pull_requests_emit_bounded_parallel_fetches_as_they_finish(
+    monkeypatch,
+) -> None:
+    """A slow repository cannot hide a completed sibling repository result."""
 
     sweep = module()
-    client = PagingClient(
+    second_observed = threading.Event()
+
+    class CompletionOrderClient(PagingClient):
+        """Hold the first repository until its completed sibling is yielded."""
+
+        def request(self, args, *, input_payload=None):
+            """Make repository completion order deterministic for the assertion."""
+
+            if args[0] == "repos/ContextualWisdomLab/first/pulls":
+                assert second_observed.wait(timeout=30)
+            return super().request(args, input_payload=input_payload)
+
+    client = CompletionOrderClient(
         {
             ("orgs/ContextualWisdomLab/repos", 1): [[
                 repository("first"),
@@ -120,17 +135,18 @@ def recording_executor(*, max_workers):
         "ThreadPoolExecutor",
         recording_executor,
     )
-    results = list(
-        sweep.list_recent_pull_requests(
-            client,
-            organization="ContextualWisdomLab",
-            repository_source="organization",
-            since="2026-08-05T00:00:00Z",
-        )
+    issues = sweep.list_recent_pull_requests(
+        client,
+        organization="ContextualWisdomLab",
+        repository_source="organization",
+        since="2026-08-05T00:00:00Z",
     )
+    first_result = next(issues)
+    second_observed.set()
+    results = [first_result, *issues]
     assert [result["repository"] for result in results] == [
-        "ContextualWisdomLab/first",
         "ContextualWisdomLab/second",
+        "ContextualWisdomLab/first",
     ]
     assert worker_limits == [2]