feat(vit): in-pipeline live orientation auto-detect

[Garrett Bischof (gwbischof)]

H5b of the #37 decomposition — the live in-pipeline orientation auto-detect (final feature PR; H6 is scripts + infra). Builds on the ptychoml pieces merged in #13.

What

On the first batch with ≥64 finite scan positions, PtychoViTInferenceOp runs ptychoml.autodetect_orientation once — a forward-physics sweep of the 8 D4 candidates against the engine's baked probe — and sets ImagePreprocessorOp.dp_orient in place. ImagePreprocessorOp buffers pre-D4 frames for the sweep (_autodetect_buf, capped 256), cleared when done. Engines without a baked probe keep the configured dp_orient (logged, sweep disabled) — so this is a no-op unless the engine was exported with --probe.

Position convention (reproduces #37 — coupled change, confirmed)

This is the half deferred from H4b; columns + ranges must move together:

• PointProcessorOp default now writes col 0 = slow axis (INENC3/y_range), col 1 = fast (INENC2/x_range) — the hxn_to_vit convention. New position_swap_xy config (default false); true reverts to the legacy fast→col0 assignment (main's old behavior).
• SaveViTResult canvas ranges swap to match: x_range_um=abs(param.y_range) (slow→width), y_range_um=abs(param.x_range).
• The sweep feeds autodetect_orientation col 0 = x (fast), so positions are swapped back via ptychoml.remap_positions(swap_xy=True) — verified identical to Enhance image processing and orientation detection in holoptycho #37's inline [[1,0]].

ptychoml reuse

autodetect_orientation + remap_positions are consumed from ptychoml — no algorithm reimplemented. (spatially_diverse_sample from #13 is used by the H5a CLI, not this path — the live buffer uses first-N-buffered frames.)

Testing

This is compose+runtime Holoscan behavior (the ptycho_holo/vit_inference smoke imports already fail on a missing TILED_BASE_URL, before compose), so it's not unit-testable in CI without fighting the infra. Safety net: it faithfully reproduces #37, the reusable bits are ptychoml-tested, the remap_positions(swap_xy=True) substitution is verified bit-identical to #37, and the column/range/swap convention is audited for internal consistency. No regression:

pixi run test   # 90 passed; the 2 vit_inference/ptycho_holo smoke failures are pre-existing.

⚠️ Note: this changes the default mosaic position mapping (col 0 = slow/y) and enables live dp_orient override when the engine has a baked probe — both intentional, matching #37. position_swap_xy=true restores the prior column convention.