fix(storage#795): align datasize output and fail-fast on flat storage_options typos

[FileSystemGuy]

Closes #795.

Summary

Reporter ran mlpstorage open training retinanet datasize for a b200/15GiB host and saw the recommendation "Number of training files: 257,173" — but the log line one second earlier said "skip_listing enabled: 1,170,301 train files", and the emitted datagen-command hint bundled num_files_train=257173 with listing_validation_interval=1000 (right bucket for 1.17M, wrong for 257K). Their later run also failed with Disallowed parameter override: storage.storage_library = s3dlio when they tried to pass the storage backend explicitly.

Three related bugs, one PR:

1. Datasize log/hint discrepancy: _apply_skip_listing_params ran in __init__ before datasize() computed its recommendation, so it read num_files_train from the workload YAML default (retinanet_b200.yaml: 1170301). Moved the call to fire in datasize() after the DRAM-derived recommendation lands in params_dict, threaded through as num_files_override. Now the INFO log, the Number of training files: result line, and the emitted datagen-hint's interval all agree.

2. Emitted datagen-hint inconsistency: same root cause — the hint's listing_validation_interval came from the pre-computed injection. Fixed by (1).

3. Flat storage.storage_library typo UX: the flat form isn't a real DLIO config path; the tool auto-injects storage.storage_options.storage_library from STORAGE_LIBRARY env in --object mode, so the user's --params had zero effect on the run. But the CLOSED disallowed-override check rejected it with a message that didn't hint at the correct spelling. Added mlpstorage_py/rules/param_hints.py with a KNOWN_PARAM_TYPOS map (three known-typo entries: storage.storage_library, storage.uri_scheme, storage.prefetch_window), a fail-fast gate in cli_parser.update_args so the user doesn't wait through MPI cluster collection just to be told the param was wrong, and a "Did you mean X?" hint appended to the runtime and submission-checker (3.6.2 CLOSED, 3.6.3 OPEN) disallowed-override messages so mlpstorage validate on stored submissions gives the same suggestion.

Backwards compatibility

• No new CLI flag. num_files_override is an internal Python parameter used only by datasize().
• Existing runs are unaffected. The datasize change is display-side only; the runtime and submission-checker file-count rules re-derive their thresholds from run metadata, not from anything the log or hint said.
• Comparability across the fix boundary is unchanged. listing_validation_interval controls startup HEAD-check density, not measured throughput; both values pre/post fix are TOOL_INJECTED_PARAMS and don't gate any rule.

Test plan

• uv run pytest tests — 2925 passed, 1 skipped, 13 deselected (+9 new)
• uv run pytest mlpstorage_py/tests — 866 passed
• uv run pytest vdb_benchmark/tests — 174 passed
• uv run pytest kv_cache_benchmark/tests — 238 passed

New tests:

• tests/unit/test_skip_listing_datasize_interval.py (3 tests) — override wins over combined_params for the 257,173 recommendation, no-override still falls back to combined_params for non-datasize commands, user-supplied --params dataset.listing_validation_interval=N still short-circuits.
• tests/unit/test_param_typo_hints.py (9 tests) — typo-map integrity, CLI-parse fail-fast for the reporter's exact command shape, multiple typos surfaced in one error, canonical nested form still accepted, runtime check appends hint, unknown-typo message unchanged.

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅