Fix checkpoint config.yaml resume merge and add --config launch flag

[eugenevinitsky]

What

• Fix the resume-time config.yaml merge in train(). It computed experiment_dir = dirname(load_model_path), but checkpoints live in <run>/models/ while config.yaml is written to <run>/, so the os.path.exists check always failed and the merge silently never ran. The block also carried a stale private copy of the architecture key list (missing all obs_norm_*/obs_range_* and waypoint keys). Both problems are fixed by deleting the block and routing train() through _merge_checkpoint_arch, which already uses the correct dirname(dirname(model_path)) layout and the canonical _ARCH_ENV_KEYS.
• Print a notice instead of silently returning when _merge_checkpoint_arch finds no config.yaml, so a fall-back to ini/CLI policy/obs layout is visible in logs.
• Add a first-class --config flag to load_config. It accepts both a run-dumped config.yaml (nested sections, including deep eval.*.* nesting) and a flat dotted cluster config (env.num_agents: 128), flattens it to dotted parser dests, validates every key against the registered argparse dests (unknown key -> ValueError naming it), and applies it via parser.set_defaults. Precedence: default.ini < drive.ini < --config < explicit CLI flags. Run-dump-only keys (git, train.use_rnn, config) are dropped.
• New tests in tests/unit_tests/test_drive_config.py covering nested yaml load + CLI-flag precedence, flat dotted yaml load, and unknown-key rejection.

Why

Best models did not reproduce: resuming or fine-tuning from a checkpoint silently rebuilt the policy and observation layout from ini defaults because the config.yaml merge never fired, and there was no way to launch a run from a saved config file at all (run dumps, cluster configs, and CLI flags were three mutually incompatible formats). Typos in a config file must fail loudly rather than no-op.

Notes

• Verification: pytest tests/unit_tests/test_drive_config.py -x -q -> 9 passed, 1 skipped; pytest tests/unit_tests -k "config or yaml" -q -> 12 passed, no regressions. End-to-end smoke: --config scripts/cluster_configs/nightly_best.yaml loads correctly through load_config (num_agents=4096, backbone_hidden_size=1024, wandb project/group, eval disables all applied). Missing-config notice verified on a nonexistent checkpoint path. ruff format --check clean; the 50 ruff check findings in pufferl.py are identical on the base commit (pre-existing).
• Behavioral change: train --load-model-path now also adopts policy_name/rnn_name (and derived train.use_rnn) plus the full _ARCH_ENV_KEYS set from the checkpoint's config.yaml — previously it adopted nothing at all due to the path bug, so any resumed run that relied on CLI flags to re-specify the architecture will now get it from the checkpoint automatically (this is the intended semantics: the net must match the weights).
• Risk: configs saved by older runs whose yaml contains keys since removed from drive.ini will now raise ValueError when passed via --config (fail-fast by design); remove the stale keys from the file to launch.
• No merge-order dependencies; single self-contained commit based on origin/3.0.

🤖 Generated with Claude Code

🤖 Generated with Claude Code