fix: avoid LAPACK deadlock and validate older JAX compatibility

[Jammy2211]

Summary

Release integration's old <0.11 override selected JAX0.10.2 and the exact-wheel replay captured all four Eigen workers blocked in LAPACK ParallelBatchMap -> CholeskyFactorization -> lapack_dpotrf_ffi. Keep the release baseline at >=0.11.2,<0.12, whose open-source build removes that path. Preserve the 115-pin incident manifest and bounded native diagnostic.

Add an independent compatibility workflow for0.9.2 and0.11.2 using the original TestPyPI wheels, Python3.12.14, exact workspace/Hands commits, three original-likelihood trials, numerical/sampler checks and the historical FFT witness. The original control/candidate comparison remains manual and preserves its inconclusive result. Package policy in companion Nerves/Hands/consumer PRs retains older JAX while excluding affected releases; this validation baseline is not a global user minimum.

API Changes

Validation dependency floor changes; diagnostic workflows and evidence added. No Python public symbols, workspace API migration, timeout increases, quarantines or readiness-rule changes. See full details below.

Test Plan

• Heart full suite:1222 passed. YAML, tenant firewall and whitespace checks pass.
• Independent Sol review CLEAN for the eight-repo compatibility change.
• Exact replay37205459198 captures native LAPACK deadlock after one pass.
• Local CPU/CUDA numerical, original-likelihood and resolver witnesses described below.
• Hosted compatibility matrix 37210342253 passes both endpoints at8721c2e: Python3.12.14,113 unchanged non-JAX pins verified, original likelihood3/3 each (0.9.2:17.433/14.117/14.184s;0.11.2:15.323/12.021/12.014s), numerical and20-iteration FFT witnesses pass. Artifacts downloaded and checked.
• Exact-head Heart unit CI 37210342215 passes Python3.12/3.13 at8721c2e.

Interleaved comparison37206724174 remains inconclusive: three controls and three candidates passed, so it does not measure failure-rate improvement. Previous-head Python3.12/3.13 CI37207278755 passed; cancelled37207278790 contributes no evidence. Original runner image and original wheel/FITS hashes were not retained.

Evidence and limits

Native capture in Heart replay37205459198 shows the 0.10.2 LAPACK/Eigen worker-pool deadlock. Tagged source lacks this path in 0.9.2 and disables it in OSS builds from 0.11.1. Only 0.10.2 was observed deadlocking; other exclusions are source-based precautions.

Original likelihood passes three CPU trials and one actual CUDA trial on each of 0.9.2/0.11.2; independent likelihood, finite-difference gradient, NUFFT, Optax and short BlackJAX execution checks pass on both backends. Historical FFT workaround is retained and its reproducer passes both versions. All 25 wheel resolver cases pass; fresh normal resolution without prereleases also selects protected synthetic stable wheels. Local Python3.12.10 differs from incident3.12.14.

Broader point-source gradient diagnostics hit the unchanged300s cap on BOTH versions after earlier checks pass; cause remains unresolved. Retained range does not certify every older version. No full release validation was rerun. Detailed receipts and compatibility CI: HeartPR280.

Development authorization

Heart remains RED: release validation FAILED (stage integrate). Live human authorization: “I authorize development investigation and repair despite the current RED reason above. Follow start-dev and record the task-specific authorization wherever required.” Compatibility plan approved: “ok yes do it properly then”. Independent Sol review CLEAN; applicable tests and targeted smoke evidence above pass. This authorizes development PRs only. Merge requires explicit human /prm; no release is authorized. Umbrella: #274.

Full API Changes (for automation & release notes)

Changed Behaviour

• Smoke/integration/notebook recipes require JAX/JAXlib>=0.11.2,<0.12.
• Compatibility diagnostics exercise older0.9.2 and baseline0.11.2 without clearing release readiness.
• Original incident comparison is manual-only; its failed control and inconclusive evidence remain preserved.

Migration

• No workspace API changes. Package release ordering and lockfile migration are documented in the companion PRs and docs/release_validation.md.

Generated by the PyAutoLabs agent workflow.

Coordinated PRs

PyAutoNerves, PyAutoHands, PyAutoFit, PyAutoGalaxy, PyAutoHeart, PyAutoArray, PyAutoLens, PyAutoCTI. Nerves must publish before repaired consumer wheels.