docs(qwen3.6): resolve the accuracy-headroom puzzle and signpost the config

Three follow-ups from review:

- The sampling note said params does not accept these knobs and then, in the
  next breath, gave the params_to_add workaround. Reworded so the omission
  reads as the deliberate scoping choice it is, not a tooling limit.
- The new "capped" definition implies ~94 recovered zeros on the QAD row is
  up to +3.5 pp of headroom; none of it appears. Say why: the sub-steps that
  stop looping still fail their tests, so the win is cost, not accuracy.
- eval_configs/scicode.yaml ships without the penalty, which now reads as an
  oversight. Note in both the README and the config that it is omitted so the
  config reproduces the published table.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Signed-off-by: Keval Morabia <28916987+kevalmorabia97@users.noreply.github.com>

examples/megatron_bridge/tutorials/Qwen3.6-35B-A3B/README.md

@@ -244,9 +244,10 @@ Sampling follows the [nvidia/Qwen3.6-35B-A3B-NVFP4](https://huggingface.co/nvidi
 > it cannot share a config with the others (`deployment.command` is global). `qwen3_coder` is what
 > the model card specifies; the template's `<tool_call>` markers resemble `hermes`, which would
 > mis-parse tool calls and silently invalidate the benchmark. Its `top_k` / `presence_penalty` go
-> through tau2's `agent_args` passthrough, which NEL's top-level `params` block does not accept —
-> that is why the other five configs omit them. To set them on an `ns_*` task, inject them through
-> the adapter's `params_to_add` instead.
+> through tau2's `agent_args` passthrough, which NEL's top-level `params` block does not accept.
+> On an `ns_*` task they go through the adapter's `params_to_add` instead — the other five configs
+> still omit them, deliberately, so they reproduce the table above; see **Output length** below
+> for when `presence_penalty` helps and when it is pure cost.
 
 > [!IMPORTANT]
 > Every task sets `num_repeats: 1`; the repeat counts above come from **launching a config that many times**. N launches give N independent `pass@1` values, which is what `mean ± sem` and the paired tests need — `num_repeats: N` instead yields a single `pass@1[avg-of-N]` with no spread. GPQA is the deliberate exception.
@@ -332,7 +333,9 @@ The obvious suspect — that `--logit_kl_topk 4096` leaves the stop tokens outsi
 | **W4A4 NVFP4 PTQ** | 32 → **1** | 39.1 → 39.1 | 1,674 → 1,476 |
 | **↳ + QAD 500 iters** | 96 → **2** | 40.2 → 39.6 | 2,340 → 1,316 |
 
-Accuracy is neutral — the deltas span −0.59 to +0.89, which is run-to-run noise in both directions — and generation gets 12-44% faster.
+Accuracy is neutral — the deltas span −0.59 to +0.89, which is run-to-run noise in both directions. Note what that implies: ~94 recovered zeros on the QAD row is up to +3.5 pp of mechanical headroom, and none of it materialised, so the sub-steps that stop looping still fail their tests. The win is cost, not accuracy — generation gets 12-44% faster.
+
+The shipped `eval_configs/scicode.yaml` deliberately leaves it out so it reproduces the results table above; add it under `adapter_config.params_to_add` to get the right-hand column.
 
 **Scope it to this failure mode, though.** On GPQA Diamond, which never caps, the same setting leaves accuracy unchanged while making generation **4.2× slower** — pure cost. Use it where runaway generation actually occurs; do not make it a global default.
 

examples/megatron_bridge/tutorials/Qwen3.6-35B-A3B/eval_configs/scicode.yaml

@@ -86,6 +86,9 @@ evaluation:
           params_to_add:
             chat_template_kwargs:
               enable_thinking: true
+            # Uncomment to suppress non-termination (see the tutorial's "Output length").
+            # Left out so this config reproduces the published results table.
+            # presence_penalty: 1.5
   tasks:
     - name: ns_scicode
       nemo_evaluator_config: