refactor: Merge cli_utils.classify_model_input into utils.model_input (#1078)

[xieofxie]

Summary

Fixes #1078.

The CLI carried two parallel -m/--model classifiers:

• utils/cli.py — enum ModelInputKind, rich ModelInput (folder metadata), and a classify_model_input that raises click.UsageError on bad input. Used by the command modules.
• utils/model_input.py — string-literal kinds, leaner ModelInput, a pure classify_model_input (returns invalid, no raising) plus resolve_model_input (Hub download). Used by the library layer (auto.py, engine.py) and inspect's hub check.

inspect even ran both on the same value. This merges them into a single, canonical classifier.

Changes

• utils/model_input.py is now the single classifier + resolver:
  • ModelInputKind is a str-based Enum, so both kind is ModelInputKind.ONNX_FILE and kind == "hub_onnx" keep working from one type.
  • ModelInput gains folder metadata (is_hf_folder, folder_has_onnx, is_winml_cli_folder) and an error message.
  • classify_model_input is pure (no I/O, never raises): it folds in hub_onnx detection, folder metadata, and the strict path / HF-id validation with friendly messages, returning kind=INVALID + error for bad input.
• utils/cli.py drops its duplicate enum/dataclass/classifier and re-exports the canonical ones. is_onnx_file_path / normalize_model_arg are unchanged.
• Command callers (build, config, export, inspect, perf) now raise click.UsageError on kind=INVALID using the classifier's error message. inspect collapses its double-classify into a single call and reports a missing local .onnx as ONNX file not found.

Notes on reconciled semantics

• A non-.onnx path-shaped value with ≥2 / segments (e.g. org/repo/path) is now INVALID ("path does not exist") rather than being forwarded to the Hub as a doomed hf_id. Library pass-through (resolve_model_input(x).local_path or x) is unaffected for these inputs.
• The classifier stays existence-independent for the .onnx kind; inspect performs the existence check to surface the friendly "ONNX file not found" message.

All tests/unit/utils and tests/unit/commands suites pass (1140 tests).

[DingmaomaoBJTU]

Approving — the classifier consolidation is a clear improvement and the core logic checks out (pure classify_model_input, str-based ModelInputKind, and the resolve_model_input(x).local_path or x library pass-through is genuinely unaffected by the org/repo/path -> INVALID change). Left a few inline notes as non-blocking follow-ups: the INVALID-guard ordering in build/perf/config (path-shaped or -c-less invalid inputs bypass the friendly guard), an inspect DRY nit, and some computed-but-unused folder-metadata fields.

[DingmaomaoBJTU]

Re-reviewed after the fixes in 928f2195e (+ the main merge and export-test hardening in 23167ab8a). Every thread from the previous pass is properly addressed and CI is green — approving.

Verified:

• build.py (was blocking): classify + INVALID guard moved up right after normalize_model_arg, so INVALID inputs are rejected before both the -c and config-less routing; the late re-classification is gone and routing reuses the up-front model_input. The config-less #553 regression is closed, and the two new tests (test_build_configless_missing_onnx_raises, test_build_configless_invalid_id_raises) exercise exactly that path.
• perf.py: path-shaped missing .onnx now gets the friendly "ONNX file not found" via an existence re-check (Path is imported at module level); comment updated to match the existence-agnostic classifier.
• inspect.py: ONNX_FILE/HUB_ONNX collapse is behavior-preserving (existing file still falls through to "not yet supported"; missing file still raises "not found").
• model_input.py: is_hf_folder/folder_has_onnx/is_winml_cli_folder dropped cleanly — no dangling references in src/ or tests/.
• tests: classifier coverage consolidated into test_model_input.py; all CI shards (commands/analyze/models/optim/remaining), lint, and CodeQL pass.