environment.yml installs PyPI onnxruntime-directml (1.24.x) but EP links pinned ORT (1.25.1) — version mismatch breaks Python EP use

[chizamd]

Summary

environment.yml installs an unpinned onnxruntime-directml (→ PyPI's latest, currently 1.24.4), but the EP links the ORT version pinned in cmake/deps.txt / CI ONNXRUNTIME_VERSION (currently 1.25.1). PyPI lags the pinned tag, so a fresh conda env create leaves the Python env with an ORT whose C-API version (v24) does not match the EP (v25). Any Python-driven EP use then fails EP registration (The requested API version [25] is not available, only [1, 24] supported) and segfaults (Windows access violation on session create).

Where it bites

• Pure-ORT Python tests (e.g. test/python/whisper/test_whisper.py, the test/python/test_*.py model suite) import onnxruntime. With the env's 1.24.4 they can't register the EP.
• The main docs/quick_start.md marks ORT build (§1) and wheel install (§4) as Optional — correct for the C++ tools (model_benchmark.exe / hip-onnx-runner.exe), which don't use the Python onnxruntime at all, but misleading for any Python usage, where a matching wheel is mandatory. Nothing in the main guide flags that distinction.
• environment.yml's unpinned line actively installs the wrong version, so even a user who later builds the correct wheel started from a mismatched env (and a user who never reaches the wheel step is silently broken for Python EP use).

Why CI doesn't catch it

CI never pip installs PyPI's onnxruntime-directml. It builds ORT 1.25.1 from source (windows-build.yml: checkout v$ONNXRUNTIME_VERSION, apply ONNXRUNTIME_PR_PATCHES, build.bat ... --build_wheel), installs that wheel on the GPU runner, and caches it. So the source-built wheel always matches the EP. The PyPI path is exercised only by humans following environment.yml + the "optional" doc steps.

Already handled (scoped to Whisper)

PR #283 / the Whisper quick-start (docs/whisper_quick_start.md §1b) now documents the mandatory source-built ORT wheel and adds a MORPHIZEN_EP_BIN override; the "ORT version must match pip" CLAUDE.md gotcha was corrected. That covers Whisper readers but not the main-flow inconsistency below.

Proposed fix (main flow — needs a maintainer decision)

1. environment.yml: either drop the unpinned onnxruntime-directml (and note that the correct ORT is installed via the source-built wheel — see quick_start §1/§4), or pin/comment it so it's clearly a placeholder that the wheel step overrides. Leaving it unpinned guarantees a wrong version on every fresh env.
2. docs/quick_start.md: keep §1/§4 "optional" for the C++-tools-only path, but add an explicit callout that for any Python EP usage the source-built ORT wheel is required and must match cmake/deps.txt (PyPI lags). Cross-link the recipe.
3. Optionally, have a single source of truth for the pinned ORT version (today it's duplicated in cmake/deps.txt and windows-build.yml); environment.yml / docs should reference it rather than hardcode.

Repro

conda env create -f environment.yml && conda activate hipdnn-ep
python -c "import onnxruntime as ort; print(ort.__version__)"   # 1.24.4 (PyPI), EP needs 1.25.1
# build EP, then run any Python EP test -> EP registration fails / access violation

Context: surfaced while running the Whisper quick-start from a freshly-cleaned workspace (PR #283 / #266).