feat(triton-cache): wire setup/teardown into entrypoint

entrypoint.py

@@ -60,6 +60,14 @@ def _early_cuda_kick() -> None:
 
 _early_cuda_kick()
 
+# Hydrate triton compilation cache from HF Hub before any triton/mamba_ssm import.
+# triton_cache_setup.py is copied next to this file by the job bash command.
+try:
+    import triton_cache_setup as _tcs
+    _tcs.setup()
+except ImportError:
+    print('[boot] triton_cache_setup not found; skipping cache hydrate', flush=True)
+
 from huggingface_hub import HfApi  # noqa: E402  (import after cuda kick)
 
 REPO_ROOT = Path('/workspace/feather')
@@ -172,6 +180,13 @@ def run_job_mode() -> int:
     print(f'[job] command={cmd}', flush=True)
     proc = subprocess.run(cmd, check=False)
 
+    # Push triton compilation cache back to HF Hub for next run.
+    try:
+        import triton_cache_setup as _tcs
+        _tcs.teardown()
+    except Exception as _tcs_err:
+        print(f'[triton_cache] teardown error (non-fatal): {_tcs_err}', flush=True)
+
     if TOKEN:
         api = HfApi(token=TOKEN)
         try: