[krafczyk]

Summary

Reduce routine Actions usage from 18 jobs per workflow to 8, and eliminate duplicate push/PR workflows:

Target	Python	Selection	Jobs
Linux	3.10–3.14	good-enough	5
Windows	3.12	good-enough	1
macOS	3.12	good-enough	1
Linux	3.12	Framework-specific heavy selections	1

• Run automatic test CI on pull requests only. Keep explicit manual suite dispatch available.
• Add workflow-level concurrency cancellation per PR. A newer eligible run cancels earlier running/pending CI for that PR; different PRs and manual dispatches use distinct groups.
• Fold the real-kernel notebook regression into Linux Python 3.12 good-enough coverage.
• Install one pinned heavy-framework stack entirely with pip. Run explicit framework/native-estimator selections rather than repeating the broad good-enough suite.
• Use XGBoost 3.4.1 consistently across extras, runtime admission, tests, and documentation.
• Explicitly copy Flax NNX candidate variables and update the JAX x64 test API for the selected stack.
• Preserve the native 97-case zero-skip gate, installed-wheel worker checks, JUnit artifacts, and explicitly requested extended suites.

Verification

• Clean isolated-worktree sdist/wheel build and 192 focused tests passed, including workflow policy, documentation, native references/adapters, NNX state isolation, runner behavior, and installed extra metadata.
• During implementation: 274 framework tests and 97 native-estimator tests passed with zero skips, including the native JUnit gate and installed-wheel worker-origin checks.
• Representative good-enough coverage was verified across initial and resumed category batches after local storage interruptions and updating the old CI-policy assertion; the remaining batch passed 1,248 ordinary tests and 2 process-sensitive cases, with expected optional-dependency skips.
• Pip dependency checks, YAML/shell/embedded-Python validation, and git diff checks passed.
• Hosted platform checks will run on this PR.

Compatibility

The exact XGBoost reader gate now selects 3.4.1. Fitted envelopes written with the former 3.4.2 candidate remain incompatible; no saved-state migration is introduced.