fix: avoid runtime compilation in regression target encoding (ENG-1093)

[jmkuebler]

Generated by Codex.

Batched regression on Torch 2.13 can invoke an eager Triton kernel and fail when no C compiler is installed. Flatten scalar targets before the v3/v3.5 column and ICL linear projections, then restore their batch shape, preserving estimator batching. This adds no startup probe or compiler detection; a fresh-process GPU test rejects compiler invocation.

Related: https://linear.app/priorlabs/issue/ENG-1093/fix-nightly-container-accuracy-failure-missing-c-compiler

Validation and limits

• RTX PRO 6000, extracted failing-image Python 3.13.15 / Torch 2.13.0+cu130 / Triton 3.7.1 runtime; same airfoil fold, v3.5-fast checkpoint and eight estimators.
• Current main (ba6ea311) fails with compiler discovery disabled. Patched source passes with a fresh Triton cache and batching enabled.
• New compiler regression test fails on unpatched main and passes with the fix. Two output/input-gradient cases cover v3 and v3.5, both target stages, and a batched non-contiguous input. The compiler-specific test runs only on CUDA with Torch 2.13+, avoiding an unnecessary Triton import on the lowest-dependency job.
• Existing v3/v3.5 architecture tests plus the initial projection checks: 195 passed. Final trimmed projection test file: 3 passed in 3.12 seconds on RTX/Torch 2.13. Pre-commit passes.
• One same-node timing comparison: constructor 24 → 23 microseconds; fit 117 → 114 ms; first predict 995 → 347 ms; warm predict 61.37 → 61.27 ms. This is a smoke comparison, not a broad performance benchmark. The baseline had a compiler available; the fix did not.
• Predictions are not bitwise identical after changing the projection kernel: mean absolute difference 0.00513, maximum 0.03156 on airfoil predictions spanning 104.9–138.8. Full accuracy-suite and Docker/HTTP validation remain outstanding.
• Logs/scripts retained at gcp-ml:/home/jonas_priorlabs_ai/strable-probe-results/container-compiler-20261001.

[github-actions]

Reapproving after a push, as this is a PR from an internal author and was already approved.