[nightshift] 20260422 multi-cleanup

[claude-nightshift]

dcebae8d — haiku
Dead code falls away
Stale comments drift with the breeze
Logs hoist at module root

Summary

Nightshift cleanup sweep across lib/iris and lib/zephyr. Two scouts produced
changes; two filed no_change with notes recorded below for follow-up.

lib/iris/src/iris (applied — 7505fd9)

• Hoisted the from iris.rpc import logging_pb2 import and the _STR_TO_ENUM
  mapping in iris/logging.py out of the str_to_log_level function body to
  module scope. The deferred import had no corresponding circular dependency
  (iris.rpc.logging_pb2 does not import iris.logging) and every file that
  imports str_to_log_level already imports logging_pb2 directly.
• Widened the str_to_log_level parameter type to str | None and folded the
  falsy-check into the function, allowing four call sites to drop their
  redundant if level_name else 0 guards.
• Fixed a stale code comment in cluster/types.py that pointed at
  iris.logging for LevelPrefixFormatter / _LEVEL_PREFIX — those helpers
  live in rigging.log_setup.

lib/zephyr/src/zephyr (applied — a3d9e7c)

• Removed the dead scatter_manifest_dir parameter from _regroup_result_refs,
  its two call sites, and the two local variables (output_stage_name,
  join_output_stage_name) that only existed to construct the now-unused path.
  The coordinator-side scatter manifest was removed in [zephyr] Eliminate coordinator-side scatter manifest consolidation #4853 but this plumbing
  was left behind.

lib/levanter/src/levanter (no_change — findings only)

Scout flagged but could not commit (sandbox denied file writes):

• Correctness bug: levanter/schedule.py:value_at_step iterates the
  schedule forward and returns the first entry whose start <= step, which
  always yields the earliest segment's value for any schedule with more than
  one ScheduleStep. The call site levanter/trainer.py:batch_axis_at_step
  depends on this, so a training run configuring train_batch_size as an
  IntSchedule gets the wrong batch size after the first segment boundary.
  Regressed in PR Reproduce Olmo 2 SFT #887 / commit 83f3f34 (renamed until → start without
  reversing the iteration order).
• utils/stat_utils.py:RunningMean.add recomputes self.total + total twice
  (harmless duplicate).
• shapes.py:conforms appears unreferenced across lib/.
• utils/background_iterable.py:80-87 has tautological except StopIteration: raise / except Exception as e: raise e blocks flagged by the AGENTS.md
  LLM-pitfall list.

lib/marin/src/marin (no_change — findings only)

Scout identified dead code but was also blocked by the sandbox:

• asdict_excluding in utils.py (~27 LOC + self-referential tests) —
  last real caller was removed when speedrun code was deleted in Remove all marin -> experiment import paths #4541.
• lib/marin/src/marin/cluster/ray.py (959 LOC, entirely unused) — already
  being removed in open PR Delete dead Marin Ray glue (cluster/ray.py + executor branches) #5028, so not duplicated here.
• evaluation/utils.py uses print instead of logger in
  download_from_gcs / upload_to_gcs.
• core/runtime.py still calls datetime.utcnow() (deprecated in 3.12).

Test plan

• ./infra/pre-commit.py --all-files --fix — clean
• uv run pytest -x lib/iris/tests/test_logging.py — 19 passed
• uv run pytest -x lib/zephyr/tests/test_execution.py — 42 passed