fix(scheduler): reject commented reviews before runner admission

.github/workflows/pr-review-merge-scheduler.yml

@@ -109,6 +109,17 @@ jobs:
       (
         github.event_name != 'repository_dispatch' ||
         github.event.client_payload.org_sweep != true
+      ) &&
+      (
+        github.event_name != 'pull_request_review' ||
+        github.event.action == 'dismissed' ||
+        (
+          github.event.action == 'submitted' &&
+          (
+            github.event.review.state == 'approved' ||
+            github.event.review.state == 'changes_requested'
+          )
+        )
       )
     runs-on: ubuntu-24.04
     # Bound scan-pr-queue to a wall-clock ceiling well short of GitHub's

CHANGELOG.md

@@ -11,6 +11,10 @@
 - Raised `hourly-review-repair.yml`'s discovery ceiling from 50 to 200 while rotating deterministic 50-PR deep-inspection windows by hourly run number. The scheduler hydrates only the selected window and stops immediately after its single dispatch, preserving access to newer PRs without quadrupling expensive review/check/comment work. See `docs/doctoring/hourly-review-repair-single-file-consolidation.md`'s 2026-09-03 follow-up.
 
 ## [Unreleased]
+- Stop `pull_request_review: submitted` events with state `commented` at the
+  merge scheduler's job-admission boundary, before a hosted runner is
+  requested. `approved`, `changes_requested`, and `dismissed` review
+  transitions retain their existing exact-PR scheduler path and permissions.
 - **Fix current-main contract drift that blocked the unscoped
   `agent-review-runtime-quality-ci.yml` "Verify scheduler and
   contextual-orchestrator review-repair contracts" step (which discovers and

docs/doctoring/pr-review-merge-scheduler-trigger-audit-20260903.md

@@ -1,5 +1,19 @@
 # Doctoring record: pr-review-merge-scheduler.yml's "fires at every step" pattern is by-design, not a bug (2026-09-03)
 
+> **2026-09-05 correction.** The broad claim below that every submitted review
+> is an actionable approval-state change was incomplete. GitHub emits
+> `pull_request_review: submitted` for `COMMENTED` reviews, which do not create
+> an `APPROVED` or `CHANGES_REQUESTED` state. On PR #1885, CodeRabbit submitted
+> `COMMENTED` reviews at 03:08:52Z, 04:27:31Z, and 05:30:37Z; the central
+> scheduler admitted runner-backed runs 33941045179, 33944606701, and
+> 33947394894 within seconds. The scheduler still needs the review trigger for
+> `APPROVED`, `CHANGES_REQUESTED`, and `dismissed`, but `COMMENTED` is now
+> rejected by the `scan-pr-queue` job-level `if` before runner acquisition.
+> The executable truth-table contract is
+> `tests/test_merge_scheduler_review_event_admission.py`. This correction does
+> not reinterpret a bot comment as formal review evidence and does not alter
+> exact-PR concurrency, review semantics, or scheduler permissions.
+
 - **Date:** 2026-09-03
 - **Subject:** the user directly observed the scheduler workflow firing repeatedly ("왜 각 모든 단계마다 Trigger
   되고 있죠?") after live evidence surfaced today of severe org-wide Actions thrashing (near-zero completion

tests/test_merge_scheduler_review_event_admission.py

@@ -0,0 +1,100 @@
+"""Executable admission contract for merge-scheduler review events."""
+
+from __future__ import annotations
+
+import ast
+import re
+from pathlib import Path
+
+import pytest
+
+
+ROOT = Path(__file__).resolve().parents[1]
+WORKFLOW = ROOT / ".github" / "workflows" / "pr-review-merge-scheduler.yml"
+
+
+def scan_job_condition() -> str:
+    """Return the normalized pre-runner condition for ``scan-pr-queue``."""
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+    scan_job = workflow.split("\n  scan-pr-queue:\n", 1)[1]
+    condition = scan_job.split("\n    runs-on:", 1)[0].split("\n    if: >-\n", 1)[1]
+    return " ".join(line.strip() for line in condition.splitlines())
+
+
+def admits_review_event(*, action: str, state: str) -> bool:
+    """Evaluate the workflow's review-event condition for one trusted fixture."""
+    expression = scan_job_condition().replace("&&", " and ").replace("||", " or ")
+    expression = re.sub(r"\btrue\b", "True", expression)
+    values = {
+        "github.event_name": "pull_request_review",
+        "github.event.action": action,
+        "github.event.review.state": state,
+        "github.event.client_payload.org_sweep": False,
+    }
+
+    def evaluate(node: ast.AST) -> object:
+        """Interpret only the boolean/comparison subset used by the job guard."""
+        if isinstance(node, ast.Expression):
+            return evaluate(node.body)
+        if isinstance(node, ast.BoolOp):
+            operands = [bool(evaluate(value)) for value in node.values]
+            return all(operands) if isinstance(node.op, ast.And) else any(operands)
+        if isinstance(node, ast.Compare) and len(node.ops) == len(node.comparators) == 1:
+            left = evaluate(node.left)
+            right = evaluate(node.comparators[0])
+            if isinstance(node.ops[0], ast.Eq):
+                return left == right
+            if isinstance(node.ops[0], ast.NotEq):
+                return left != right
+        if isinstance(node, ast.Constant):
+            return node.value
+        if isinstance(node, (ast.Attribute, ast.Name)):
+            key = ast.unparse(node)
+            if key in values:
+                return values[key]
+        raise AssertionError(f"unsupported scheduler expression node: {ast.dump(node)}")
+
+    return bool(evaluate(ast.parse(expression, mode="eval")))
+
+
+@pytest.mark.parametrize(
+    ("action", "state", "expected"),
+    [
+        ("submitted", "commented", False),
+        ("submitted", "approved", True),
+        ("submitted", "changes_requested", True),
+        ("dismissed", "commented", True),
+    ],
+)
+def test_review_event_admission_truth_table(
+    action: str, state: str, expected: bool
+) -> None:
+    """Admit only review transitions that can change merge eligibility."""
+    assert admits_review_event(action=action, state=state) is expected
+
+
+def test_review_filter_preserves_exact_pr_group_and_least_privilege() -> None:
+    """Filtering COMMENTED reviews must not weaken scheduler trust boundaries."""
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+    assert (
+        "github.event_name == 'pull_request_review' && "
+        "format('pr-{0}', github.event.pull_request.number)" in workflow
+    )
+    assert (
+        "cancel-in-progress: ${{ github.event_name == 'pull_request_target' || "
+        "github.event_name == 'pull_request_review' || "
+        "github.event_name == 'repository_dispatch' }}" in workflow
+    )
+    assert "permissions:\n  contents: read" in workflow
+
+    scan_header = workflow.split("\n  scan-pr-queue:\n", 1)[1].split(
+        "\n    env:\n", 1
+    )[0]
+    for permission in (
+        "actions: write",
+        "checks: read",
+        "contents: write",
+        "id-token: write",
+        "pull-requests: write",
+    ):
+        assert permission in scan_header