Ship v0.1.1–v0.1.5: first-live-run fixes and dashboard viewer

- v0.1.1: config persistence — FileHelper paths made absolute
  (/media/startos/volumes/main/...); relative paths resolved into the JS
  runtime's ephemeral cwd so action saves never reached /data
- v0.1.2: preJobStopContainers (Configure Grading) — docker-stop resident
  vLLM containers on the head Spark at job start, no auto-restart
- v0.1.3: preflight auth (LiteLLM master_key gates /models),
  poll-until-loaded, crash fast-fail (restarting counts as dead)
- v0.1.4: HF_HUB_OFFLINE/TRANSFORMERS_OFFLINE in airgapped serving
  (--internal network has no DNS); grader _post timeout 600→1800s for
  ~3.6 tok/s GB10 generation
- v0.1.5: dashboard viewer survives the periodic background refresh;
  download buttons for deck reports, deck JSON, and SCORECARD.md
- .gitignore: .startos/ build workspace, start-technologies/

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

orchestrator/serving.py

@@ -65,6 +65,22 @@ def remove_network(cfg: dict, log) -> None:
     sc.run(head, f"docker network rm {shlex.quote(net_name(cfg))} 2>/dev/null; true", timeout=30)
 
 
+# ---------------------------------------------------------- resident models
+def clear_resident_containers(cfg: dict, log) -> None:
+    """Stop the user-listed containers on the head Spark so a grading job gets
+    the GPU to itself (both Sparks normally run an always-on 31B vLLM).
+
+    Deliberately NOT restarted after the job: whatever owns them is responsible
+    for bringing them back (the Gazette's Fleet job reloads its own models)."""
+    names = (cfg.get("preJobStopContainers") or "").replace(",", " ").split()
+    if not names:
+        return
+    head = sc.head(cfg)
+    quoted = " ".join(shlex.quote(n) for n in names)
+    log(f"[serving] freeing the GPU on {head.host}: docker stop {' '.join(names)}")
+    sc.run(head, f"docker stop {quoted} 2>/dev/null; true", timeout=180)
+
+
 # ------------------------------------------------------------------ vLLM
 def _vllm_run(cfg: dict, model: dict, hf_token: str | None) -> tuple[sc.Spark, str]:
     """Build the docker run command for one model on its assigned Spark."""
@@ -75,6 +91,10 @@ def _vllm_run(cfg: dict, model: dict, hf_token: str | None) -> tuple[sc.Spark, s
     parser = cfg.get("toolCallParser", "hermes")
     tools = (f"--enable-auto-tool-choice --tool-call-parser {shlex.quote(parser)} " if parser else "")
     env = f"-e HF_TOKEN={shlex.quote(hf_token)} " if hf_token else ""
+    if cfg.get("networkMode") == "airgapped":
+        # The --internal network has no DNS/egress; without offline mode the HF
+        # hub client dies on name resolution even with the model fully cached.
+        env += "-e HF_HUB_OFFLINE=1 -e TRANSFORMERS_OFFLINE=1 -e VLLM_NO_USAGE_STATS=1 -e DO_NOT_TRACK=1 "
     cache = _hf_cache(cfg)
     name = _vllm_name(alias)