[tdk4-dev]

Add an experimental Nuke Python Group that bakes a frame range with the separately installed public CorridorKey 2 runtime, either on the local CUDA workstation or over SSH. Completed RGBA EXRs are verified and connected through Read/Premult nodes; no PyTorch dependency is loaded into Nuke.

The adapter includes source/custom ranges, optional alpha hints, cancellation and timeouts, per-user worker serialization, checksummed uploads/downloads, interrupted-download recovery, and explicit scene-linear Rec.709/PAR handling. It keeps each bake in a separate job directory and does not redistribute model code, weights, or commercial OFX binaries.

Validation:

• 25 dependency-light contract/process tests passed, including real synthetic EXR roundtrips, HDR/negative RGB, alpha, PAR, missing/corrupt frames, archive safety, cancellation/timeouts, and RNG/hash repeatability.
• Full-repository Ruff 0.15.4 lint and formatting checks passed.
• Native Linux Nuke 15.1v4 tests passed for Group creation, nested exports, alpha-hint extraction, premultiplication, Qt completion callbacks and changed-settings handling.
• A bounded three-frame 640x440 PAR=2 GPU bake passed locally and through SSH on an RTX 4080, including downstream native Nuke rendering and retrieval without re-running inference.

Qualification remains incomplete: local-from-Nuke and SSH inference gave different pixels on identical exported/prepared inputs (maximum alpha difference approximately 0.028). Fixed seeds and deterministic PyTorch settings did not remove that difference. Both modes function, but numerical parity is unresolved; this PR stays a draft and is not a production-readiness claim. Full-resolution shots, other GPUs and Nuke versions remain unqualified.

Targets public CK2 runtime commit 485f5e9ab759ead6138cc990776184b59d6ea971. Half/quarter resolution uses explicit adapter-side down/up-sampling because that revision's optimized EXR path rejects non-native model_scale.