Two bugs surfaced when running `run_eval.py --build-only` against the EP matrix on a CPU-only host:

1. Every combo for the 'no native EP' subset (mlas/dml/openvino) was reported as `[FAIL @ complete]` even though export/optimize/quantize/model.onnx all landed correctly. `_run_build` was funnelling build-only results through `_extract_onnx_path`, which scans stdout for a `Final artifact:` marker that `winml build --no-compile` never prints, and falls back to the global cache which build-only doesn't populate (`-o <dir>` writes elsewhere). In build-only mode there is no downstream consumer of the path, so trust exit-code 0 directly and record `build_out` to keep the per-component bookkeeping balanced.

2. QNN/VitisAI combos failed at the optimize stage with `Requested EP 'qnn' is not available on this system`. `_run_optimize_stage` calls `resolve_device(device, ep=ep)` purely to pick the right `has_rule_data_for_ep` key for the progress bar, but that helper raises when the EP isn't installed locally -- even when the rest of the pipeline (export + optimize + quantize) runs on CPU and the EP is only needed at compile time. Soft-fail the lookup *only when* `config.compile is None` (i.e. `--no-compile` or a config that explicitly opts out); otherwise re-raise so configs that will compile still fail fast here instead of deep inside the compile stage.

Also moves `--clean-cache` from per-combo to per-model in `_run_build_only`: combos for the same model share the same HF download, so clearing between combos forced N redundant re-downloads of the same weights.

…facts feed

Running --build-only over the 8-EP matrix for many models fills local disk.
Two additions keep disk bounded:

1. Export dedup: the export.onnx stage is EP/device-independent, so every
   combo produces an identical export. After each combo builds, its export is
   hash-compared against a per-model canonical: the first is moved to
   <model_dir>/_shared/, later identical ones are deleted. One export copy on
   disk instead of 8.

2. --upload: after a model's combos are built, publish the model dir to the
   Modelkit Azure Artifacts feed as a Universal Package version, then delete it
   locally. Auth via az login (no PAT); the azure-devops extension is ensured
   and login verified up front (aborts otherwise so disk isn't silently filled).
   Version is 0.0.0-<run-stamp>-<model-slug> (valid SemVer 2.0; date stamp
   groups a batch). --continue + --run-stamp resume an interrupted batch from
   the build_only_uploads.json manifest without rebuilding uploaded models;
   --keep-local, --upload-skip-existing, and feed/package args round it out.

- _hash_files: stop hashing unreadable files to a fixed sentinel; propagate
  OSError and have _dedup_export keep the export in place instead of risking
  deletion of an artifact never verified identical.
- _is_publish_conflict: narrow detection to specific version-exists / HTTP 409
  markers (drop bare 'conflict'/'409') so an unrelated message can't trigger
  exists-skipped and rmtree the local model dir.
- build.py _run_optimize_stage: narrow the no-compile EP fallback to only
  swallow EP-not-available ValueErrors; re-raise malformed device/EP names.
- Warn when --continue is used with --build-only but without --upload (no
  local-disk resume exists, so everything is rebuilt).
- Document that the pinned-EP auto-device path delegates precision to winml
  config's auto-detection.
- Fix misleading --upload-skip-existing help: it does not skip the build.

When a mid-run upload failed because Azure CLI was unavailable (not logged in, token expired, or az hung and was killed), the model was marked 'failed' and kept locally while the batch continued. Every subsequent model hit the same az failure and its local copy piled up, filling the disk (a single 7B LLM matrix is ~450 GB).

Add _is_az_unavailable() to distinguish a host-level az/login problem from a per-package publish error (network blip or version conflict), and abort the run immediately (exit 3) when an upload fails for that reason. The user re-runs 'az login' and resumes with --continue + the same --run-stamp; already-uploaded models are skipped.

Uploading all 8 EP/device combos as one Universal Package version timed out on
large models and left the local artifacts behind, filling the disk. Each combo
is now built, uploaded, and deleted on its own.

- Version is per combo: 0.0.0-<run-stamp>-<ep>-<device>-<model-slug>, so each
  package is small (lower timeout risk) and can be retried/resumed on its own.
- Local artifacts are removed after every outcome (uploaded, version-exists,
  timeout, upload-failed, build-failed) unless --keep-local, so disk stays
  bounded. A timeout/failure is recorded and the run continues; only a
  host-level az failure (not logged in / token expired) aborts.
- Every (model, combo) outcome is written to build_only_results.json (replaces
  build_only_uploads.json) for auditing, with or without --upload; it also
  drives per-combo --continue resume.
- Export dedup now applies only without --upload (each uploaded combo is
  self-contained).

Adds _classify_upload helper and unit tests for the version format, results
I/O, az-unavailable classification, and the per-combo cleanup/timeout/abort
orchestration.

…atrix

# Conflicts:
#	src/winml/modelkit/commands/build.py

CodeQL flagged the best-effort 'remove now-empty model dir' cleanup in _run_build_only as an empty 'except OSError: pass'. Add an explanatory comment (no behaviour change) so the intent -- ignoring a non-empty/locked dir -- is documented.