docs: fill CLI reference gaps for data prep and rollout collection

[lbliii]

Summary

Fills documented gaps in the CLI reference and installation pages. All CLI additions were verified against nemo_gym/prompt.py and nemo_gym/rollout_collection.py.

• ng_materialize_prompts — full parameter table, example, and a note clarifying when to use it vs. ng_prepare_data. Closes docs: add missing CLI commands to reference and clarify when to use each #1347.
• resume_from_cache — params-table entries for ng_collect_rollouts and ng_e2e_collect_rollouts, plus a Resume Interrupted Runs section covering the *_materialized_inputs.jsonl artifact, per-rollout incremental flush, (task_index, rollout_index) matching, fallback behavior, and the stale-cache footgun. Closes docs: document resume_from_cache for rollout collection #1239.
• Generation parameters — dedicated section documenting that temperature, top_p, and max_output_tokens are passed via ++responses_create_params.<field> rather than as standalone flags. Documents shallow (top-level) merge behavior for nested overrides. Closes bug: could not find temperature CLI params due to docs segmented control #637.
• PyPI install default — adds a PyPI tab as the first install option in installation.mdx. Closes docs: swap install path to pip install nemo-gym with 0.3.0 PyPI release #1191.
• v0.3.0 mirror — mirrors CLI reference and installation changes into stable docs.

Review feedback addressed

• Copilot review: clarified that responses_create_params overrides use a shallow merge (dict | overrides in _preprocess_rows_from_config), so nested Hydra overrides replace the entire nested dict rather than deep-merging fields.

Verification

• npm run check (fern check) passes with 0 errors.

---

Note: Replaces #1498. That PR's branch (docs/cli-reference-cleanup) could not be rewritten due to branch protection (no force-push/delete); an unsigned commit (91526fce) left DCO failing. This PR uses a single signed commit on lbliii/docs-cli-dco-fix.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

🌿 Preview your docs: https://nvidia-preview-lbliii-docs-cli-dco-fix.docs.buildwithfern.com/nemo/gym

Here are the markdown pages you've updated:

• nemo/gym/get-started/installation
• nemo/gym/reference/cli-commands
• nemo/gym/get-started/installation
• nemo/gym/reference/cli-commands

[marta-sd]

@lbliii In the meantime I refactored CLI in PR #1630 and updated the docs, thus the conflicts. Here's the updated version of the content you prepared: #1741 Note that it should be used for latest / 0.4.0, but not for 0.3.0 (this is new CLI syntax, not available in 0.3.0)

[sephmard]

Good additions overall — the Resume Interrupted Runs section and the render vs. collate Note are both genuinely useful. Two things to resolve before merge: the latest/ Generation Parameters example still uses old ng_collect_rollouts syntax (tracked in #1741 — suggest absorbing those fixes here before merging), and the about shallow merge reads better adjacent to the gym eval run section rather than at the top of the page before the commands are introduced.

[sephmard]

The PyPI tab shows pip install nemo-gym with no Python version constraint, then uv pip install nemo-gym with --python 3.12. The plain pip path could silently use the wrong Python version. Worth either adding a note ("requires Python 3.12") or showing pip with a version check first.

[sephmard]

Nit: Generation Parameters Tip placement.

The about shallow merge is placed at the very top of the page in the quick-reference section (line 95 context), before any commands are defined. A reader scanning top-to-bottom hits this tip before they know what gym eval run or responses_create_params are. It reads better adjacent to the gym eval run section where it's acted on.

[lbliii]

Updated this PR to cover the remaining review feedback and absorb the relevant latest-doc content from #1741.

Summary:

• Replaced bare PyPI pip install examples with Python 3.12-explicit commands in both main/latest and v0.3.0 installation docs.
• Moved the shallow-merge guidance out of the early Hydra overview and into the gym eval run generation parameters section, next to the flags where readers use it.
• Added the latest/main generation-parameters example using current gym eval run syntax: --temperature, --top-p, --max-output-tokens, plus raw ++responses_create_params... overrides.
• Absorbed the remaining latest-only docs: update content added in PR 1675 to match the new CLI syntax #1741 data-prep update by moving gym dataset collate before gym dataset render and expanding the render contract around responses_create_params.input.
• Left v0.3.0 on the release-appropriate CLI syntax, per the note that the new syntax is latest/0.4.0-only.

Validation:

• npm run check passes locally with 0 errors.
• PR checks are passing on head 1c66b32f; the remaining blocker is review approval.