v0.15.8

0.15.8 hardens upgrade/config surfaces and serializes bulk watcher root-spec
intake behind the v1 one-open-closure-target policy.

Added

• Added millrace --version and millrace version for package-version
  visibility in installed and module-entrypoint environments.
• Added millrace run daemon --monitor-log PATH to write basic monitor output
  to a file while keeping stdout monitor mode independently selectable.
• Added first-class Codex model_reasoning_effort config on
  runners.codex and per-stage stages.<stage>.model_reasoning_effort, with
  compile, request, runner artifact, and run-inspection visibility.

Changed

• Bulk watcher intake now respects the v1 one-open-closure-target policy as
  queue backpressure: unrelated root specs remain queued while the active root
  lineage runs or reaches Arbiter, status reports the deferred-root count, and
  direct stale-state recovery preserves the open closure target.
• stages.<stage> and runners.codex.permission_by_stage now accept learning
  stages such as professor.
• Workspace doctor now resolves the selected mode against workspace-local
  deployed assets, matching runtime compile/startup behavior for custom modes.
• millrace upgrade can intentionally localize removed managed assets through
  --localize-removed or --localize-removed-from, and runtime asset
  manifests now ignore cache artifacts such as __pycache__ and *.pyc.
• Runtime shutdown now clears process_running, status suppresses stale
  process_running truth when no active ownership lock exists, and
  clear-stale-state reports applied when it only clears that stale process
  bit.