Add analysis scripts and experiment configurations for bridge attention and sensitivity studies

- Introduced `bridge_attention_ceiling_check.py` for variance decomposition analysis on bridge attention configurations.
- Added `bridge_attention_readout.py` to perform per-tower gate and contribution readouts, including AUC sanity checks.
- Created multiple JSON configuration files for backbone replication experiments, including anonymous CV variants and basic backbones.
- Implemented sensitivity experiments to evaluate the impact of axial length inclusion and EfficientNetV2-M performance at higher resolutions.
- Added a memory probe script to assess GPU memory usage during training with EfficientNetV2-M.

v4/distributed/client.py

@@ -168,20 +168,28 @@ def _gpu_info() -> str:
 # rsync helpers
 # ──────────────────────────────────────────────────────────────
 
-def _rsync(src: str, dst: str, delete: bool = False):
+def _rsync(src: str, dst: str, delete: bool = False,
+           excludes: list[str] | None = None):
     cmd = ["rsync", "-az", "--info=progress2"]
     if delete:
         cmd.append("--delete")
+    for pat in (excludes or []):
+        cmd.append(f"--exclude={pat}")
     cmd += [src, dst]
     subprocess.run(cmd, check=True)
 
 
 def _sync_code(server_ssh: str, server_path: str, local_path: str):
-    """Pull v4/ source from server → local (overwrites local changes)."""
+    """Pull v4/ source from server → local (overwrites local changes).
+
+    results/ is excluded so the client never overwrites or deletes its own
+    per-job output directory, and so it never pulls down the full corpus of
+    historical results from the server.
+    """
     src = f"{server_ssh}:{server_path}/v4/"
     dst = f"{local_path}/v4/"
     Path(dst).mkdir(parents=True, exist_ok=True)
-    _rsync(src, dst, delete=True)
+    _rsync(src, dst, delete=True, excludes=["results/"])
 
 
 def _upload_results(server_ssh: str, server_path: str, local_path: str,

v4/distributed/server.py

@@ -413,8 +413,33 @@ def submit_job(job: JobSubmit):
         )
         if cur.rowcount == 0:
             existing = conn.execute(
-                "SELECT job_id FROM jobs WHERE args=?", (args_json,)
+                "SELECT job_id, state FROM jobs WHERE args=?", (args_json,)
             ).fetchone()
+            # If the existing duplicate is a terminal failure, drop it and
+            # take the new submission — saves an explicit /jobs/clear round-
+            # trip when re-deploying after a fix.
+            if existing["state"] == "failed":
+                conn.execute(
+                    "DELETE FROM jobs WHERE job_id=?", (existing["job_id"],)
+                )
+                conn.execute(
+                    "INSERT INTO jobs "
+                    "(job_id, run_name, module, args, output_dir, priority, created_at) "
+                    "VALUES (?,?,?,?,?,?,?)",
+                    (job_id, job.run_name, job.module, args_json,
+                     job.output_dir, job.priority, _now()),
+                )
+                print(
+                    f"[server] requeued failed {existing['job_id']} → {job_id} "
+                    f"({job.run_name})",
+                    flush=True,
+                )
+                return {
+                    "job_id":          job_id,
+                    "duplicate":       False,
+                    "requeued":        True,
+                    "previous_job_id": existing["job_id"],
+                }
             job_id = existing["job_id"]
             print(f"[server] duplicate ignored ({job.run_name}) → {job_id}", flush=True)
             return {"job_id": job_id, "duplicate": True}