Fail fast on non-finite RL metrics

[ISEEKYAN]

Summary

• add a CPU-only streaming gate for non-finite ppo_kl, loss, and grad_norm metrics
• exit with status 42 on the first NaN or infinity while preserving the rejected line
• expose k3-rl-finite-gate and document the pipefail/tee launcher integration

Verification

• focused finite-gate tests: 12 passed
• repository Ruff format check: 58 files already formatted
• repository Ruff lint check: passed
• diff whitespace check: passed

The tests include an actual Bash pipefail pipeline and verify that a non-finite metric propagates exit status 42 through tee.

[ISEEKYAN]

Closed per repo owner's standing freeze on new k3 PRs (2026-08-01): k3 has a single open delivery PR (#20) and all further changes must land as commits on its branch feat/k3-three-stage-rl-clean. The change itself (fail-fast on non-finite RL metrics) is accepted and will be re-landed there — closing this PR is a process correction, not a rejection of the fix.