submit: keep eval-gpu image repo as a fixed literal

revert the env-var indirection; the audit sweep stays prose-only and
must not alter dispatch resolution. the repo id and job namespace are
real deployed identities, kept as documented constants.

submit.py

@@ -20,8 +20,8 @@ poll of an HF Jobs GPU eval + boot-time stuck-pending sweep. The
 handler validates the upload, uploads the zip to
 ``submissions/<id>.zip``, appends a ``status: pending`` row to
 ``results.jsonl`` (under a process-wide lock), spawns a daemon thread
-that dispatches a per-submission HF Job against the eval-gpu
-Docker Space image and polls
+that dispatches a per-submission HF Job against the
+``HuggingAI4Engineering/cadgenbench-eval-gpu`` image and polls
 ``inspect_job`` until the job's stage is terminal. On COMPLETED the
 worker downloads ``reports/<id>.json`` (the Job already uploaded
 ``reports/<id>.{html,json}`` to the submissions dataset), reads
@@ -111,7 +111,7 @@ from huggingface_hub import (
 )
 from huggingface_hub.errors import EntryNotFoundError
 
-from leaderboard import HF_DATA_REPO, HF_ORG, HF_SUBMISSIONS_REPO
+from leaderboard import HF_DATA_REPO, HF_SUBMISSIONS_REPO
 
 logger = logging.getLogger(__name__)
 
@@ -136,16 +136,11 @@ STUCK_PENDING_REASON = "evaluation interrupted by Space restart"
 BOOT_SWEEP_ENV = "CADGENBENCH_DISABLE_BOOT_SWEEP"
 
 # HF Jobs target. The eval-gpu image is hosted as a Docker Space
-# (paused; image-only) under the org; the repo id is derived from
-# HF_ORG so an org rename is one Space-variable change. Jobs run
-# under the personal `michaelr27` namespace, which is required and
-# stays a fixed constant: dispatch is billed to that account
-# (no-bill for HF employees per Round 6 of space-setup/leandro.md),
-# so it is deliberately not configurable. a10g-large fits
-# cadgenbench evaluate --workers 8 comfortably in 46 GB RAM.
-EVAL_GPU_SPACE = os.getenv(
-    "CADGENBENCH_EVAL_GPU_SPACE", f"{HF_ORG}/cadgenbench-eval-gpu"
-)
+# (paused; image-only) at HuggingAI4Engineering/cadgenbench-eval-gpu.
+# Jobs run under the personal `michaelr27` namespace (no-bill for
+# HF employees per Round 6 of space-setup/leandro.md). a10g-large
+# fits cadgenbench evaluate --workers 8 comfortably in 46 GB RAM.
+EVAL_GPU_SPACE = "HuggingAI4Engineering/cadgenbench-eval-gpu"
 EVAL_JOB_FLAVOR = "a10g-large"
 EVAL_JOB_NAMESPACE = "michaelr27"
 EVAL_JOB_TIMEOUT = "30m"