refactor: introduce WinMLManifest class and rename to winml_manifest.json

[xieofxie]

Summary

Renames build_manifest.json → winml_manifest.json to align with the existing winml_build_config.json naming convention, introduces a WinMLManifest dataclass as the single source of truth for the manifest schema, and wires winml export to also generate the manifest.

Changes

New: utils/manifest.py

• ManifestStage — dataclass for one pipeline stage entry (name, status, filename, timing, optional quant metrics).
• WinMLManifest — full manifest dataclass with:
  • to_dict() / save() for serialisation (drops None fields, injects schema_version)
  • load() / from_dict() for deserialisation (forward-compatible — unknown keys land in extras)
  • find(directory) to discover all *winml_manifest.json in a directory
  • manifest_path_for(output_dir, prefix) for canonical path computation

Producers wired to WinMLManifest

Producer	Before	After
build/hf.py	raw dict → json.dumps	WinMLManifest → .save()
build/onnx.py	raw dict → json.dumps	WinMLManifest → .save()
commands/export.py	(no manifest)	WinMLManifest → .save()

Filename rename

All references to build_manifest.json updated to winml_manifest.json across source, tests, and docs (15 files).

Export now generates manifest

winml export writes winml_manifest.json alongside the exported ONNX with source: "export", stage timing, and HTPExporter stats. Composite exports get per-component prefixed manifests (e.g. model_decoder_winml_manifest.json).

Notes

• Consumers (inference/engine.py, inspect/resolver.py, serve/app.py) still read raw dicts via json.loads — they can adopt WinMLManifest.load() / .from_dict() incrementally.
• The on-disk JSON format is backward-compatible: same schema_version: 1, same field names.