Add EXQ-008/009 experiments + runner --loop mode for continuous operation

New experiment scripts (smoke-tested):
- experiments/candidate_count_ablation.py (EXQ-008, MECH-063)
  Tests whether E3 trajectory selection over multiple candidates is load-bearing.
  Conditions: SINGLE(n=1), STANDARD(n=32), EXPANDED(n=128).
- experiments/residue_weight_sensitivity.py (EXQ-009, MECH-062)
  Tests whether rho_residue routing weight is load-bearing.
  Conditions: RHO_ZERO(0.0), RHO_LOW(0.1), RHO_STANDARD(0.5), RHO_HIGH(2.0).

Runner improvements:
- experiment_runner.py: --loop / --loop-interval flags for continuous operation.
  After queue exhaustion, polls experiment_queue.json every N seconds (default 60)
  for newly-added items. Lets the runner stay alive between sessions.
- Initial queue display now prunes already-completed items on startup.

Queue updates (experiment_queue.json):
- EXQ-008 (MECH-063, high priority)
- EXQ-009 (MECH-062, high priority)
- EXQ-010..013 (low priority, extended-seed re-runs of MECH-059/056/061/060 PASS
  results for confidence accumulation — keeps runner busy between new experiments)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: nooarche

experiment_queue.json

@@ -107,6 +107,102 @@
       "estimated_runtime_min": 22,
       "status": "pending",
       "status_reason": "Script written and smoke-tested 2026-02-27. Addresses EXQ-004 FAIL root cause: atomic GridWorld has no multi-step action structure for completion gate to protect. MacroStepGrid with MACRO_LENGTH=4."
+    },
+    {
+      "queue_id": "EXQ-008",
+      "backlog_id": "EVB-0036",
+      "claim_id": "MECH-063",
+      "title": "E3 Candidate Count Ablation",
+      "description": "Tests whether the number of E2 trajectory candidates available to E3 is load-bearing for harm avoidance. Three conditions: SINGLE (1 candidate, no real selection), STANDARD (32, default), EXPANDED (128, wider search). PASS: STANDARD < SINGLE (selection matters) AND EXPANDED <= STANDARD*1.10 (no collapse at high counts).",
+      "script": "experiments/candidate_count_ablation.py",
+      "args": ["--episodes", "200", "--seeds", "7", "42", "99"],
+      "seeds": 3,
+      "conditions": 3,
+      "episodes_per_run": 200,
+      "estimated_runtime_min": 22,
+      "priority": "high",
+      "status": "pending",
+      "status_reason": "Script written 2026-02-28. Tests MECH-063 (trajectory_selection.candidate_count_load_bearing). Uses standard GridWorld with num_candidates override on generate_trajectories()."
+    },
+    {
+      "queue_id": "EXQ-009",
+      "backlog_id": "EVB-0032",
+      "claim_id": "MECH-062",
+      "title": "Residue Routing Weight Sensitivity",
+      "description": "Tests whether the E3 residue routing weight (rho_residue) is load-bearing. Four conditions: RHO_ZERO (rho=0, agent is ethically amnesic), RHO_LOW (0.1), RHO_STANDARD (0.5, default), RHO_HIGH (2.0). PASS: RHO_STANDARD < RHO_ZERO (residue routing helps) AND RHO_HIGH <= RHO_ZERO*1.20 (high weight doesn't catastrophically destabilise).",
+      "script": "experiments/residue_weight_sensitivity.py",
+      "args": ["--episodes", "200", "--seeds", "7", "42", "99"],
+      "seeds": 3,
+      "conditions": 4,
+      "episodes_per_run": 200,
+      "estimated_runtime_min": 30,
+      "priority": "high",
+      "status": "pending",
+      "status_reason": "Script written 2026-02-28. Tests MECH-062 (residue.routing_weight_load_bearing). Sets config.e3.rho_residue per condition. Key claim: residue field must have positive routing weight to influence trajectory selection."
+    },
+    {
+      "queue_id": "EXQ-010",
+      "backlog_id": "EVB-0037",
+      "claim_id": "MECH-059",
+      "title": "Control Plane Precision Separation (extended seeds)",
+      "description": "Extended-seed re-run of EXQ-000 (MECH-059 PASS). Adds seeds 13, 77, 123, 256, 314, 999 for tighter confidence intervals. Same MERGED vs SEPARATED conditions. Contributes to P1 evidence accumulation for the precision channel separation claim.",
+      "script": "experiments/control_plane_precision_separation.py",
+      "args": ["--episodes", "200", "--seeds", "13", "77", "123", "256", "314", "999"],
+      "seeds": 6,
+      "conditions": 2,
+      "episodes_per_run": 200,
+      "estimated_runtime_min": 30,
+      "priority": "low",
+      "status": "pending",
+      "status_reason": "Low-priority re-run for confidence accumulation. EXQ-000 PASS (seeds 7,42,99). Adds 6 more seeds."
+    },
+    {
+      "queue_id": "EXQ-011",
+      "backlog_id": "EVB-0039",
+      "claim_id": "MECH-056",
+      "title": "Residue Trajectory Placement (extended seeds)",
+      "description": "Extended-seed re-run of EXQ-001 (MECH-056 PASS). Adds seeds 13, 77, 123, 256, 314, 999. TRAJECTORY-WIDE vs ENDPOINT-ONLY conditions. Builds replication depth for the residue placement claim.",
+      "script": "experiments/residue_trajectory_placement.py",
+      "args": ["--episodes", "200", "--seeds", "13", "77", "123", "256", "314", "999"],
+      "seeds": 6,
+      "conditions": 2,
+      "episodes_per_run": 200,
+      "estimated_runtime_min": 30,
+      "priority": "low",
+      "status": "pending",
+      "status_reason": "Low-priority re-run for confidence accumulation. EXQ-001 PASS (seeds 7,42,99). Adds 6 more seeds."
+    },
+    {
+      "queue_id": "EXQ-012",
+      "backlog_id": "EVB-0041",
+      "claim_id": "MECH-061",
+      "title": "Commitment Boundary Token Reclassification (extended seeds)",
+      "description": "Extended-seed re-run of EXQ-003 (MECH-061 PASS). Adds seeds 13, 77, 123, 256, 314, 999. PRE_COMMIT vs POST_COMMIT channel correlation conditions. Tightens confidence on commit boundary separation.",
+      "script": "experiments/commitment_boundary_validation.py",
+      "args": ["--episodes", "200", "--seeds", "13", "77", "123", "256", "314", "999"],
+      "seeds": 6,
+      "conditions": 2,
+      "episodes_per_run": 200,
+      "estimated_runtime_min": 30,
+      "priority": "low",
+      "status": "pending",
+      "status_reason": "Low-priority re-run for confidence accumulation. EXQ-003 PASS (seeds 7,42,99). Adds 6 more seeds."
+    },
+    {
+      "queue_id": "EXQ-013",
+      "backlog_id": "EVB-0043",
+      "claim_id": "MECH-060",
+      "title": "Write-Locus Contamination (extended seeds)",
+      "description": "Extended-seed re-run of EXQ-005 (MECH-060 PASS). Adds seeds 13, 77, 123, 256, 314, 999. FULL vs CONTAMINATED_DURABLE vs CONTAMINATED_RESIDUE. Builds replication depth on the write-locus separation result.",
+      "script": "experiments/write_locus_contamination.py",
+      "args": ["--episodes", "200", "--seeds", "13", "77", "123", "256", "314", "999"],
+      "seeds": 6,
+      "conditions": 3,
+      "episodes_per_run": 200,
+      "estimated_runtime_min": 45,
+      "priority": "low",
+      "status": "pending",
+      "status_reason": "Low-priority re-run for confidence accumulation. EXQ-005 PASS (seeds 7,42,99). Adds 6 more seeds."
     }
   ]
 }

experiment_runner.py

@@ -296,6 +296,23 @@ def main():
         action="store_true",
         help="Show what would run without executing"
     )
+    parser.add_argument(
+        "--loop",
+        action="store_true",
+        help=(
+            "After exhausting the queue, poll experiment_queue.json every "
+            "--loop-interval seconds for new items instead of exiting. "
+            "Lets the runner stay alive so clicking Start in the explorer "
+            "keeps it running without a manual restart."
+        ),
+    )
+    parser.add_argument(
+        "--loop-interval",
+        type=int,
+        default=60,
+        metavar="SECONDS",
+        help="Seconds between queue re-checks in --loop mode (default: 60)",
+    )
     args = parser.parse_args()
 
     status_path = args.status_file or find_default_status_path()
@@ -358,54 +375,104 @@ def handle_signal(sig, frame):
         return
 
     print(f"[runner] PID {os.getpid()} — ready to run {len(items)} queued experiments", flush=True)
+    if args.loop:
+        print(f"[runner] Loop mode: will poll queue every {args.loop_interval}s after exhaustion", flush=True)
 
     completed_ids = {c["queue_id"] for c in existing_completed}
 
-    for item in items:
-        queue_id = item["queue_id"]
-
-        # Skip already completed
-        if queue_id in completed_ids:
-            print(f"[runner] Skipping {queue_id} ({item['claim_id']}) — already completed", flush=True)
-            continue
-
-        # Skip items with no script
-        script = REPO_ROOT / item["script"]
-        if not script.exists():
-            print(f"[runner] Skipping {queue_id} ({item['claim_id']}) — script not found: {item['script']}", flush=True)
-            # Update queue item status in live status file
-            for qi in status["queue"]:
-                if qi["queue_id"] == queue_id:
-                    qi["status"] = "needs_script"
-            write_status(status, status_path)
-            continue
+    # Prune any already-completed items from the initial queue display
+    status["queue"] = [qi for qi in status["queue"] if qi["queue_id"] not in completed_ids]
+    write_status(status, status_path)
 
-        # Run it
-        result = run_experiment(item, status, status_path, calibration)
+    while True:
+        ran_any = False
 
-        # Move to completed
-        completed_entry = {
-            "queue_id": queue_id,
-            "backlog_id": item.get("backlog_id", ""),
-            "claim_id": item.get("claim_id", ""),
-            "title": item.get("title", ""),
-            "description": item.get("description", ""),
-            "result": result["result"],
-            "result_summary": result["result_summary"],
-            "completed_at": result["completed_at"],
-            "output_file": result.get("output_file", ""),
-        }
-        status["completed"].append(completed_entry)
-        completed_ids.add(queue_id)
+        for item in items:
+            queue_id = item["queue_id"]
 
-        # Remove from queue display
-        status["queue"] = [qi for qi in status["queue"] if qi["queue_id"] != queue_id]
-        status["current"] = None
+            # Skip already completed
+            if queue_id in completed_ids:
+                continue
+
+            # Skip items with no script
+            script = REPO_ROOT / item["script"]
+            if not script.exists():
+                print(f"[runner] Skipping {queue_id} ({item['claim_id']}) — script not found: {item['script']}", flush=True)
+                # Update queue item status in live status file
+                for qi in status["queue"]:
+                    if qi["queue_id"] == queue_id:
+                        qi["status"] = "needs_script"
+                write_status(status, status_path)
+                continue
+
+            # Run it
+            result = run_experiment(item, status, status_path, calibration)
+            ran_any = True
+
+            # Move to completed
+            completed_entry = {
+                "queue_id": queue_id,
+                "backlog_id": item.get("backlog_id", ""),
+                "claim_id": item.get("claim_id", ""),
+                "title": item.get("title", ""),
+                "description": item.get("description", ""),
+                "result": result["result"],
+                "result_summary": result["result_summary"],
+                "completed_at": result["completed_at"],
+                "output_file": result.get("output_file", ""),
+            }
+            status["completed"].append(completed_entry)
+            completed_ids.add(queue_id)
+
+            # Remove from queue display
+            status["queue"] = [qi for qi in status["queue"] if qi["queue_id"] != queue_id]
+            status["current"] = None
 
+            write_status(status, status_path)
+            print(f"[runner] Done: {queue_id} — {result['result']}", flush=True)
+
+        # Queue pass complete
+        if not args.loop:
+            break
+
+        # Loop mode: wait, then reload queue for any newly-added items
+        status["idle"] = True
+        status["current"] = None
         write_status(status, status_path)
-        print(f"[runner] Done: {queue_id} — {result['result']}", flush=True)
+        if ran_any:
+            print(f"[runner] Pass complete. Waiting {args.loop_interval}s before re-checking queue…", flush=True)
+        else:
+            print(f"[runner] No new items. Waiting {args.loop_interval}s…", flush=True)
+
+        time.sleep(args.loop_interval)
+
+        # Reload queue — picks up any new items added while we slept
+        queue_data = load_queue()
+        calibration = queue_data.get("calibration", {})
+        items = queue_data.get("items", [])
+
+        new_pending = [i for i in items if i["queue_id"] not in completed_ids]
+        if new_pending:
+            print(f"[runner] Found {len(new_pending)} new item(s): "
+                  f"{[i['queue_id'] for i in new_pending]}", flush=True)
+            # Rebuild queue display in status
+            new_queue_display = []
+            for i in new_pending:
+                new_queue_display.append({
+                    "queue_id": i["queue_id"],
+                    "backlog_id": i.get("backlog_id", ""),
+                    "claim_id": i.get("claim_id", ""),
+                    "title": i.get("title", ""),
+                    "description": i.get("description", ""),
+                    "estimated_minutes": round(estimate_minutes(i, calibration), 1),
+                    "status": "pending",
+                    "status_reason": i.get("status_reason", ""),
+                })
+            status["queue"] = new_queue_display
+            status["idle"] = False
+            write_status(status, status_path)
 
-    # All done
+    # All done (non-loop path)
     status["idle"] = True
     status["current"] = None
     status["runner_pid"] = None