Pin GDPVal evaluator launcher to 0.2.6

plugins/modelopt/skills/evaluation/SKILL.md

@@ -70,7 +70,9 @@ GDPVal:
    unsandboxed. Verify with `gdpval-sif.sh --check`. `.env` needs `HF_TOKEN`, `INFERENCE_API_KEY`, `TAVILY_API_KEY`,
    `INFERENCE_JUDGE_URL`, `GDPVAL_SIF_DIR`, and `NEMO_EVALUATOR_TRUST_PRE_CMD=1` (the
    config has a `pre_cmd`). Thinking mode is mandatory (non-thinking loses ~86%).
-4. Dry-run → launch. **`limit_samples` is inert on the gym path** (the gym runs all
+4. Run both dry-run and launch through `"$SKILL_DIR/scripts/nel-gdpval.sh"`; it pins
+   the required 0.2.6 launcher and avoids older launchers' unset
+   `NEL_INVOCATION_ID` failure. **`limit_samples` is inert on the gym path** (the gym runs all
    220 tasks regardless), so there is no cheap canary: watch the real run's first
    ~20–30 min for the SIF-sandbox line and judge auth, and cancel if wrong. See the
    recipe's Canary section.

plugins/modelopt/skills/evaluation/recipes/examples/gym_gdpval/example_gym_gdpval.yaml

@@ -144,6 +144,8 @@ evaluation:
     # deliverables are NOT auto-uploaded — they stay on disk for inspection.
     # Drop "_cache" if you WANT them uploaded as artifacts.
     PERSIST_DELIVERABLES_DIR: lit:/results/gdpval/deliverables_cache
+    # Requires launcher 0.2.6 via $SKILL_DIR/scripts/nel-gdpval.sh. That launcher
+    # assigns this stable ID before the runtime re-export (including across resumes).
     NEL_INVOCATION_ID: runtime:NEL_INVOCATION_ID
   # Installs apptainer + squashfuse into the eval container (needs
   # NEMO_EVALUATOR_TRUST_PRE_CMD=1 in the launching shell). See references/gym-gdpval.md.

plugins/modelopt/skills/evaluation/recipes/tasks/aa_gym/gdpval.md

@@ -15,6 +15,11 @@ in the suite — **220 tasks**, `num_repeats=1`, 4 judge trials per rollout.
 It runs on the **0.2.6 `nel` launcher** as a `nemo_gym` task (NOT nel-next), so
 Steps 1–9 apply — but with the branch differences below.
 
+Run it through `"$SKILL_DIR/scripts/nel-gdpval.sh"`, which pins the launcher to
+0.2.6. Do not use an unversioned `nel` from PATH: older launchers fail before client
+startup when the config forwards `NEL_INVOCATION_ID`. The shared reference explains
+the failure signature and dry-run check.
+
 ## What makes GDPVal different (not a normal `aa/` task)
 
 - **Standalone** — one gym eval per config. Never add GDPVal to a multi-task

plugins/modelopt/skills/evaluation/references/gym-gdpval.md

@@ -9,6 +9,22 @@ run **inline in the eval container** (`install_on_the_fly`) via `ng_prepare_benc
 is `recipes/examples/gym_gdpval/` and the per-task pointer is
 `recipes/tasks/aa_gym/gdpval.md`.
 
+Always invoke GDPVal through the pinned wrapper, even if `nel` is already on PATH:
+
+```bash
+"$SKILL_DIR/scripts/nel-gdpval.sh" --version  # must report nemo_evaluator_launcher: 0.2.6
+"$SKILL_DIR/scripts/nel-gdpval.sh" run --config <gdpval-config.yaml> --dry-run
+"$SKILL_DIR/scripts/nel-gdpval.sh" run --config <gdpval-config.yaml>
+```
+
+This is a correctness requirement, not only a dependency pin. Launcher 0.2.6 writes
+the generated `NEL_INVOCATION_ID` into `run.sub` before environment-variable
+re-exports. Older launchers can emit `export NEL_INVOCATION_ID="${NEL_INVOCATION_ID}"`
+without first assigning it, then exit with `NEL_INVOCATION_ID: unbound variable`
+under `set -u` before the evaluation client starts. In a dry-run, verify that a
+literal assignment appears before the re-export; do not substitute `SLURM_JOB_ID`,
+because it changes across the benchmark's walltime-resume chain.
+
 ## Where each piece runs
 
 | Component | Where |
@@ -184,7 +200,7 @@ mount source, and `raise ValueError` listing the missing ones **before** any
 | `GDPVAL_REF_FILES_DIR` | lit:/gdpval_ref_files | shared-FS ref-file staging (node-local /tmp breaks multi-node Ray) |
 | `PERSIST_DELIVERABLES_DIR` | lit | where deliverables persist (see MLflow note) |
 | `GDPVAL_MAX_TURNS` | lit (optional) | Stirrup turn cap (default 100; golden uses 250) |
-| `NEL_INVOCATION_ID` | runtime | run id |
+| `NEL_INVOCATION_ID` | runtime | stable run id assigned by launcher 0.2.6; do not use `SLURM_JOB_ID` |
 
 `INFERENCE_JUDGE_URL` is the judge host — config (from `.env`), substituted as the
 literal `<INFERENCE_JUDGE_URL>` placeholder in `gdpval_judge.base_url`, **not**

plugins/modelopt/skills/evaluation/scripts/nel-gdpval.sh

@@ -0,0 +1,37 @@
+#!/usr/bin/env bash
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# GDPVal requires launcher 0.2.6. In particular, 0.2.6 assigns the generated
+# NEL_INVOCATION_ID in run.sub before evaluation.env_vars re-exports it. Older
+# launchers can instead fail under `set -u` before the evaluation client starts.
+set -euo pipefail
+
+NEL_GDPVAL_SPEC="${NEL_GDPVAL_SPEC:-nemo-evaluator-launcher[all]==0.2.6}"
+
+case "${1:-}" in
+  -h|--help)
+    echo "usage: nel-gdpval.sh <nel arguments>"
+    echo "       nel-gdpval.sh --version"
+    exit 0
+    ;;
+esac
+
+command -v uvx >/dev/null 2>&1 || {
+  echo "ERROR: 'uvx' is required to run the pinned GDPVal launcher" >&2
+  exit 1
+}
+
+exec uvx --python 3.10 --from "$NEL_GDPVAL_SPEC" nel "$@"