fix: skip shot ranking during cut rejection checks

docs/pipeline-setup.md

@@ -137,6 +137,8 @@ New runs retain a source SHA-256 in `run.json`. Stages refuse missing or changed
 before launching; existing outputs stay on disk. Legacy manifests without a digest need a new
 submission instead of blindly reusing cached state. Orchestrated runs reject cuts instead of
 silently trimming into another run directory. The direct runner retains its explicit shot controls.
+Reject-only cut checks skip shot ranking and its model loads; their detector arguments match
+the cache key. Cut detection and rejection still run before any stage.
 
 `wander step` admits one unresolved attempt per step in a run directory. A filesystem lock
 serializes admission, then releases before execution so independent steps can proceed. Detached

docs/pipeline-validation.md

@@ -104,6 +104,20 @@ sign-in; inherited unrelated MCP sign-in and duplicate-hook warnings did not pre
 turn. The broken system CLI was then updated to that version and its exec/resume startup
 checks passed. Evidence: `.context/evidence/worker-reply-resume/`.
 
+A further bottle delivery check used the real Codex agent to inspect the retained inputs,
+launch one detached `package_people` stage, end its turn, and resume when packaging finished.
+The stage completed in 78.3 seconds; the agent read its result and requested viewer review.
+The live admin showed one successful attempt and distinguished five saved-input stages from
+fresh execution. All 240 animated person frames, room, object assets, original audio, and
+presentation matched the prior delivery. Camera/source URLs correctly named the new run.
+Browser checks loaded both people, enabled the original soundtrack, advanced playback, reached
+frame 119 at 9.93 seconds, wrapped to frame zero, and moved beyond the room without an edge
+fade. No console errors were observed. The supervisor stopped and the run was paused.
+This was the explicitly requested fifth packaging execution after four successful delivery
+checks, with no automatic retry, provider commands, paid-ledger change, or quality acceptance.
+It joins actual agent control to retained packaging and viewer delivery; fresh reconstruction
+and headset behavior remain unverified. Evidence: `.context/evidence/bottle-agent-delivery/`.
+
 Regression tests cover questions arriving before stage/agent launch, blocked watchdog resumes,
 retained transcript/session evidence, and answers arriving between a turn's final journal read
 and the supervisor's next pass. Quiet turns do not repeatedly wake on their own status notes.

scripts/run_clip.py

@@ -3077,8 +3077,8 @@ def valid_report(value):
         "--threshold",
         str(a.cut_threshold),
     ]
-    if a.shot is not None or a.all_shots:
-        cmd.append("--no-score")  # the choice is already made; do not pay to rank
+    if no_score:
+        cmd.append("--no-score")  # chosen shots and reject-only checks do not need ranking
     say("cut check: looking for cuts before anything is spent")
     try:
         p = subprocess.run(

scripts/tests/test_run_clip_cut_cache.py

@@ -61,13 +61,14 @@ def restore(self):
         run_clip.ROOT = self.old_root
         self.temp.cleanup()
 
-    def args(self, threshold=0.1, shot=None, all_shots=False):
+    def args(self, threshold=0.1, shot=None, all_shots=False, reject_cuts=False):
         return SimpleNamespace(
             clip=str(self.clip),
             name="same-run",
             cut_threshold=threshold,
             shot=shot,
             all_shots=all_shots,
+            reject_cuts=reject_cuts,
         )
 
     def cache_path(self):
@@ -133,6 +134,18 @@ def test_score_mode_change_same_threshold_misses(self):
         self.assertEqual(len(self.calls), 2)
         self.assertIn("--no-score", self.calls[-1])
 
+    def test_reject_cuts_skips_ranking_and_matches_its_cached_detector_args(self):
+        with patch.object(run_clip.subprocess, "run", side_effect=self.fake):
+            run_clip.cut_check(self.args())
+            result = run_clip.cut_check(self.args(reject_cuts=True))
+            cached = run_clip.cut_check(self.args(reject_cuts=True))
+        self.assertEqual(len(self.calls), 2)
+        self.assertNotIn("--no-score", self.calls[0])
+        self.assertIn("--no-score", self.calls[1])
+        self.assertIn("--no-score", result["_cacheKey"]["detectorArgs"])
+        self.assertFalse(result["continuous"], "skipping ranking must retain cut detection")
+        self.assertEqual(result, cached)
+
     def test_detector_key_change_misses(self):
         with patch.object(run_clip.subprocess, "run", side_effect=self.fake):
             run_clip.cut_check(self.args())