fix(logging): retarget file handler transactionally

[TZZheng]

Summary

• mark and reuse only LingTai-owned logging handlers while preserving handlers installed by embedding applications
• serialize the complete handler discovery and mutation transaction with a shared reentrant lock
• retarget file logging transactionally: fully open a replacement before swapping it in and closing the old handler
• preserve the working file handler when replacement opening fails, allowing a later retry
• cover ownership, repeatability, reconfiguration, concurrency, and failure recovery with focused regressions

Validation

Exact candidate: 51511659e6fe820c25e12046751d64806e037406

• PYTHONPATH=src python -m pytest -q tests/test_kernel_logging.py — 6 passed
• git diff --check canonical/main...HEAD — clean
• merge-tree check — clean
• worktree status — clean

Fresh Claude/Codex review reports are still pending; this body makes no review-verdict claim.

Risk

The process-local lock serializes rare setup calls. Unmarked external handlers remain untouched. Retargeting replaces the owned file-handler object only after its new stream opens successfully, avoiding failed-open corruption without adding rollback machinery.

Fixes #749