Add a rule file for the python launch and model scripts

[Zhichenzzz]

Stacked on #2356 (base is yueming/script-rewrite, review that first). Docs only — no runtime code changes.

#2356 finishes moving scripts/ to python, but the convention that produced it lives only in that PR's description and in the tests. This adds .claude/rules/launch-and-model-scripts.md so the next launcher follows it without archaeology, and registers it in AGENTS.md next to general-code-style.md.

Every rule is taken from the post-#2356 code or from a test that already enforces it, not from preference:

• Python, not shell. load_model_args resolves scripts/models/<megatron_model_type>.py by file name, so a shell model definition cannot be found at all — and test_shell_script_hygiene.py already rejects a source scripts/models/x.sh in a docker patch. The surviving examples/**/*.sh are named as legacy awaiting a port, so nobody copies one as a template.
• Model definitions. File name equals the megatron_model_type; one pure model_args(**kwargs) -> str; variants derive via load_sibling_model_args / moe_layer_freq; any environment knob a model script reads has to be in CLEARED_ENV, which test_every_environment_knob_a_model_script_reads_is_frozen enforces.
• Launcher anatomy. Discovery is scripts/**/run_*.py, so the naming rule is what buys snapshot coverage. Public module-level functions are entrypoints (prepare, execute) and the suite asserts each issues commands, so helpers must be _-prefixed. Shell access goes through command_utils; the self-executing launchers under examples/ are called out as legacy pinned by test_self_executing_launchers.py.
• Nothing hardcoded about the machine. --model-dir / --data-dir / --output-dir instead of literals, U.repo_base_dir instead of a checkout path, wandb only via get_default_wandb_args, environment read at call time (the num_nodes default_factory), --num-gpus-per-node always passed, --rollout-num-gpus dropped under --colocate, an already-joined cluster expressed with MILES_SCRIPT_EXTERNAL_RAY=1.
• One launcher per recipe family. Merge near-duplicates behind a frozen _Recipe table keyed by a Literal --model-name; keep separate files when the recipes are different experiments, with run_glm45_355b_a32b_8node.py as the worked example and a stated threshold for when not to merge.
• Snapshot coverage is the review artifact. Regenerate with MILES_UPDATE_LAUNCH_SCRIPT_SNAPSHOTS=1 pytest tests/manual/launch_scripts and read the diff; launchers must import and run without GPU, checkpoints or network; don't grow the harness denylists without a test that fails once the entry stops being true.
• Docs follow the launcher. Adding/renaming/deleting a launcher updates the pages that invoke it, including hf download --local-dir destinations.

Testing

pre-commit run --files .claude/rules/launch-and-model-scripts.md AGENTS.md clean. No code paths touched.