fix(logging): never delete .log files held by live processes

[yishuiliunian]

Summary

• cleanup_old_logs was unlinking still-open log files of long-lived hub / agent-server processes, stranding KB of logs in unreadable write-only fds and making classifier failures undebuggable.
• Two-part fix: open new log fds as O_RDWR (so an unlinked inode can still be re-opened via /dev/fd/N), and teach cleanup to skip any .log whose filename PID is still alive.
• Investigation context: hub-only sessions ran --decision=classifier; users reported repeated classifier failures with no extractable log evidence. Root cause was that each ephemeral sub-agent's startup re-ran cleanup and deleted the hub/server's older-mtime log.

Changes

• src/log_writer.rs
  • RotatingFileWriter::new and rotation path: OpenOptions adds .read(true).
  • Refactor cleanup_old_logs to delegate to cleanup_with_alive_filter(impl Fn(u32) -> bool); production wiring passes bootstrap::is_alive.
  • New helper pid_from_log_filename parses both loopal-{ts}-{pid}.log and rotated *.N.log shapes.
  • 3 inline tests covering filename parsing, alive-PID protection under count overflow, and the new read-capable fd.
• src/bootstrap/mod.rs
  • pub(crate) use discovery::is_alive so log_writer can reach it.

Test plan

• bazel build //:loopal passes
• bazel build //:loopal --config=clippy zero warnings
• bazel build //:loopal --config=rustfmt clean
• bazel test //:loopal-unit-test — 3 new tests pass alongside 40 existing
• CI green