test(e2e): winml inspect E2E tests + fix --list-tasks crash

[DingmaomaoBJTU]

Summary

Adds tests/e2e/test_inspect_e2e.py — 39 tests covering winml inspect.
Offline tests run without downloading any model weights; network tests hit
HuggingFace Hub with real model IDs.

Also fixes a crash in get_known_tasks(): HF_MODEL_CLASS_MAPPING contains
(model_type, None) sentinel entries for per-model-type defaults; iterating
without a None guard caused sorted() to raise
TypeError: '<' not supported between instances of 'NoneType' and 'str'
whenever --list-tasks was invoked.

uv run pytest tests/e2e/ -m "e2e and not network" -k inspect_e2e   # offline (30)
uv run pytest tests/e2e/ -m "e2e and network"     -k inspect_e2e   # network (9)

CLI commands under test

# ── offline: CLI surface ─────────────────────────────────────────────
winml inspect
winml inspect --help
winml inspect --model-type bert --format xml

# ── offline: --list-tasks ────────────────────────────────────────────
winml inspect --list-tasks

# ── offline: --model-type / --model-class (no weight download) ───────
winml inspect --model-type bert -f json
winml inspect --model-type bert -t feature-extraction -f json
winml inspect --model-type resnet -f json
winml inspect --model-class BertForMaskedLM -f json
winml inspect --model-type bert --verbose -f json
winml inspect --model-type bert --hierarchy -f json
winml inspect --model-type totally_nonexistent_model_xyz_123

# ── network: real model IDs ──────────────────────────────────────────
winml inspect -m bert-base-uncased -f json
winml inspect -m bert-base-uncased -f json -t feature-extraction
winml inspect -m bert-base-uncased -f json -t next-sentence-prediction
winml inspect -m microsoft/resnet-50 -f json
winml inspect -m facebook/convnext-tiny-224 -f json
winml inspect -m google/vit-base-patch16-224 -f json
winml inspect -m openai/clip-vit-base-patch32 -f json
winml inspect -m openai/clip-vit-base-patch32 -f json -t image-feature-extraction
winml inspect -m facebook/detr-resnet-50 -f json

Coverage

TestInspectCliSurface — CLI surface (11 tests, offline)

• winml inspect (no args) → exit 2 with UsageError
• winml inspect --help → exit 0
• All flags documented in --help: -m, -f, --model-type, --model-class, --list-tasks, -v/--verbose, -H/--hierarchy
• --format xml → exit non-zero, error mentions the bad choice

TestInspectListTasks — --list-tasks (6 tests, offline)

• Exit 0, non-empty output, every line is a non-empty string
• Includes well-known tasks: feature-extraction, mask-generation
• Works without any -m / --model-type / --model-class argument
• Contrast: bare winml inspect → exit 2; winml inspect --list-tasks → exit 0

TestInspectModelTypeOnly — offline inspection (13 tests, offline)

• --model-type bert and --model-type resnet resolve correct model_type + task in JSON
• --model-class BertForMaskedLM resolves to bert / fill-mask
• All 14 keys in EXPECTED_TOP_KEYS present in JSON output
• loader, exporter, winml sub-sections have required fields
• exporter.input_tensors and output_tensors are lists
• -v / --verbose and table format both exit 0
• --hierarchy without -m → hierarchy: null (no crash)
• Unknown model type exits non-zero

TestInspectBert — bert-base-uncased (3 tests, network)

• Auto-detect resolves to fill-mask, task_source == "TasksManager"
• Explicit feature-extraction override accepted
• next-sentence-prediction: clean success or clean error (no raw traceback)

TestInspectVision — vision models (3 tests, network)

• microsoft/resnet-50, facebook/convnext-tiny-224, google/vit-base-patch16-224
• Auto-detect → image-classification for all three

TestInspectCLIP — openai/clip-vit-base-patch32 (2 tests, network)

• Auto-detect → feature-extraction or zero-shot-image-classification
• Explicit image-feature-extraction override accepted

TestInspectDETR — facebook/detr-resnet-50 (1 test, network)

• Auto-detect → object-detection

[xieofxie]

Code review — 3 issues, posted inline so each can be resolved individually. The get_known_tasks() None-guard fix looks clean.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[timenick]

LGTM. All three review comments cleanly addressed in 41d1947:

• 8 test_help_documents_* collapsed into a single parametrized test with a class-scoped help_output fixture (one --help invocation per class)
• test_invalid_format_* merged into one method
• _run extracted to tests/_helpers.py with the correct Result return type, shared by both the CLI and E2E test files

The get_known_tasks() None-sentinel fix is also correct — verified that MODEL_CLASS_MAPPING (e.g. in models/hf/sam.py) intentionally uses (model_type, None) entries as per-model-type defaults.

🤖 Generated with GitHub Copilot CLI