feat(vit): derive ViT inner_crop from the ONNX probe geometry

[Garrett Bischof (gwbischof)]

H4a of the #37 decomposition — derive the ViT-patch inner_crop from the model's baked probe instead of always auto-deriving it. (H4b, true vit-only mode, follows separately.)

Changes

New holoptycho/engine_probe.py:

• find_onnx_for_engine — map a compiled {name}_v{ver}.engine to its source onnx/{name}/{ver}/*.onnx (the model_manager cache convention).
• inner_crop_from_onnx — load the ONNX, extract the complex probe (probe_real/probe_imag initializers, or two 2-D constant graph outputs), and hand it to ptychoml.inner_crop_from_probe (the inscribed-square geometry, added in feat(preprocess): add inner_crop_from_probe — ViT-patch crop from probe geometry ptychoml#12). Falls back to None when onnx is unavailable or the probe can't be extracted.

Wired in PtychoApp.compose(): inner_crop precedence is explicit config > ONNX-probe-derived > None (then SaveViTResult auto-derives min(H,W)//4).

Why a separate module

The helpers are pure (Path/re/onnx/ptychoml — no Holoscan/Tiled), and vit_inference doesn't use them internally (only compose() does). Putting them in engine_probe.py keeps them out of vit_inference's import-time Tiled setup so they're unit-testable.

Tests

tests/test_engine_probe.py (7) fully covers find_onnx_for_engine — the holoptycho-specific filename logic (match, first-sorted, multi-digit version + underscored name, bad name, missing dir, empty dir) — plus the graceful-None fallback. The inscribed-square geometry is tested in ptychoml; the ONNX extraction path runs in real model loads (onnx isn't in the CI test env, so inner_crop_from_onnx returns None there).

pixi run test   # 85 passed; the 2 vit_inference/ptycho_holo smoke failures are
                # pre-existing (module load needs a reachable TILED_BASE_URL).

Depends on

NSLS2/ptychoml#12 (merged); pixi.lock bumped.