[Wave 4] Low/info — Module sweep: Trainer core, OOM recovery, auto-revert & GRPO rewards (part 4/4) #711

Wave 4 · 6 finding(s) in 6 fix package(s). Fix these whenever a pull request touches this module ("clean the module you touch").

Findings

• CR-A02-141 (code, Info) — Convention rules with no guard show drift: bare NOSONAR, DEBUG-only BLE001 handlers, pragma: no cover on defensive branches, library log-and-raise, typing style
  severity Info · reachability unreachable · intent intended-but-harmful · confidence likely · single reviewer · reproduced end to end
  Failure scenario. model.py:638: a failed expert recast is visible only at DEBUG, so an operator expecting MoE VRAM savings gets none and no warning. The other items are convention hygiene.
  Where. forgelm/trainer.py:1693-1708, forgelm/model.py:638, forgelm/compliance.py:1005, forgelm/compliance.py:2321
  Recommended fix. Add AST guards for rule-coded NOSONAR, BLE001 handler log level and pragma placement. Fix the DEBUG-only handlers first.

• CR-A02-142 (code, Info) — The eval_steps/save_steps divisibility rule is correct but is never checked against a real transformers TrainingArguments
  severity Info · reachability default · intent intended-and-acceptable · confidence likely · independently confirmed · reproduced end to end · already tracked: PARTIALLY-FIXED CORE-14 — docs/roadmap/phase-16-trust-surface-hardening.md:292 (S2 done); CHANGELOG.md:21-25
  Failure scenario. A transformers or trl upgrade changes the save/eval constraint, for example requiring the multiple also under eval_strategy='epoch', or relaxing it. ForgeLM's rule and its tests stay green and the operator hits the late exit-2 (or a false exit-1) the check was added to prevent.
  Where. tests/test_trainer.py:1377-1440, forgelm/trainer.py:483-495, tests/runtime_smoke.py:1-109, tests/test_trainer_sft_config.py:69
  Recommended fix. Add the differential test described above, one per supported trainer type, using the real SFTConfig, DPOConfig and so on with use_cpu=True and report_to='none'. The existing 'cadence mismatch passes --dry-run' item is unaffected.

• CR-A02-143 (code, Info) — ConfigError is raised from the ForgeTrainer constructor, not train(); train() re-raises arbitrary exceptions
  severity Info · reachability default · intent contract-violation · confidence likely · independently confirmed · not reproduced (code trace)
  Failure scenario. A caller wraps only train() in try/except ConfigError and misses constructor failures. Other failure types propagate unmapped.
  Where. docs/reference/library_api_reference.md:264, docs/reference/library_api_reference-tr.md:264, forgelm/trainer.py:366-371, forgelm/trainer.py:1693-1709
  Recommended fix. Move ConfigError to the constructor row. State that train() re-raises the underlying exception after emitting pipeline.failed.

• CR-A02-144 (code, Info) — model.reverted records the regression only as free text; no structured delta or threshold fields
  severity Info · reachability default · intent intended-and-acceptable · confidence likely · independently confirmed · reproduced end to end
  Failure scenario. An auditor or dashboard parsing model.reverted to decide whether a revert was a real regression or threshold tuning has to regex free text, or join to a different event.
  Where. forgelm/trainer.py:992, forgelm/trainer.py:1406-1409, forgelm/trainer.py:1460-1464, forgelm/trainer.py:1489-1492
  Recommended fix. Emit measured, threshold and delta as structured fields on model.reverted, or amend RTP R-07 to point at the gate *.evaluation_completed events.

• CR-A02-145 (code, Info) — model.reverted reason vocabulary in the docs does not match the emitters
  severity Info · reachability undocumented · intent contract-violation · confidence likely · independently confirmed · reproduced (plumbing only)
  Failure scenario. An audit consumer matching on the documented reason set misses this revert.
  Where. forgelm/trainer.py:863 (confirmed), forgelm/trainer.py:978-979, docs/usermanuals/en/evaluation/auto-revert.md:102 (confirmed), docs/usermanuals/tr/evaluation/auto-revert.md:102
  Recommended fix. Add invalid_threshold to auto-revert.md (EN+TR) and the docstring.

• CR-A02-146 (code, Info) — save_final_model before train() reports the missing trainer as a save fallback, raises an unexplained AttributeError, and leaves an empty final_model directory
  severity Info · reachability undocumented · intent intended-but-harmful · confidence likely · independently confirmed · reproduced end to end
  Failure scenario. A caller saves adapters before or without training (for example a resume or export helper). They get a misleading 'fallback' warning, a bare AttributeError and an empty final_model/ that later blocks forgelm approve.
  Where. forgelm/trainer.py:1711-1749, forgelm/trainer.py:1713, forgelm/trainer.py:1720-1729, forgelm/trainer.py:1735
  Recommended fix. Check the precondition before makedirs: if the trainer has not been built, raise a clear RuntimeError, or save from self.model. Keep the missing-attribute case out of the narrow fallback catch.

Done when

• Every finding above is fixed or closed with a reason (tick it here).
• Code fixes carry a test; doc fixes update the Turkish mirror in the same PR.

_{Found by the October 2026 ForgeLM code and documentation review against commit cab2947. Code links point at that commit; line numbers may have moved since. Finding IDs (CR-…, DOC-…, TECH-…) are stable references for this backlog.}