[Wave 4] Low/info — Module sweep: Training orchestration, multi-stage pipelines, resume & no-train modes (part 3/3) #714

Wave 4 · 9 finding(s) in 9 fix package(s). Fix these whenever a pull request touches this module ("clean the module you touch").

Findings

• CR-A04-083 (code, Low) — User manual's event table states pipeline.started/completed fire unconditionally; the code and webhook_schema.md skip them on --resume-from and on an approval gate
  severity Low · reachability default · intent contract-violation · confidence confirmed · independently confirmed · reproduced end to end
  Failure scenario. Resume: after stage 2 failed, --resume-from stage2 sent training.start, training.success, then pipeline.completed(completed) for the same pl_, with no pipeline.started (POST lines 17-19). Gated: stage 2 with require_human_approval exits 4, and the last event is approval.required, with no pipeline.completed (lines 34-38). A receiver built from the user manual treats both as faults or open runs.
  Where. docs/usermanuals/en/operations/webhooks.md:47-48, docs/usermanuals/tr/operations/webhooks.md:47-48, docs/reference/webhook_schema.md:52-57, forgelm/cli/_pipeline.py:1131-1139
  Recommended fix. Copy the 'fresh runs only' and 'not on gated_pending_approval' qualifiers into the user manual table (EN+TR); drop 'single-stage' from webhook_schema.md:35.

• CR-A04-084 (code, Low) — A checkpoint whose trainer_state.json is present but optimizer.pt (or scheduler.pt) is missing resumes at exit 0 with a fresh optimizer and LR schedule, and nothing logs it
  severity Low · reachability documented-opt-in · intent intended-but-harmful · confidence likely · single reviewer · reproduced end to end
  Failure scenario. An operator restores checkpoint-N from a bucket where optimizer.pt failed to upload. --resume continues silently with a reset optimizer and a warm-restarted LR (a 1e-4 spike instead of a near-zero LR), and the run reports success.
  Where. forgelm/cli/_resume.py:13-66 (the cited forgelm/trainer.py:1231 was not re-verified), forgelm/trainer.py:1231
  Recommended fix. As part of the resume completeness check, when save_only_model is off, require optimizer.pt (or the sharded or DeepSpeed equivalent) and scheduler.pt. If they are missing, fail or at least WARN, and record the downgrade.

• CR-A04-085 (code, Info) — Pipeline orchestrator repeats config merges and state/manifest rewrites per transition (measured negligible)
  severity Info · reachability default · intent intended-and-acceptable · confidence likely · single reviewer · reproduced with stubs
  Failure scenario. None that is user-visible. A synthetic 30-stage pipeline with a stubbed trainer made 60 merge_pipeline_stage_config calls, 61 _save_state calls and 62 audit events (each fsynced), and the whole run took 0.11 s. A real stage costs minutes to hours.
  Where. forgelm/cli/_pipeline.py:988, forgelm/cli/_pipeline.py:1083, forgelm/cli/_pipeline.py:1370, forgelm/cli/_pipeline.py:1504
  Recommended fix. No action needed. Optionally cache the _scan_output_dirs merge results for reuse by the stage loop. Expected gain is negligible (milliseconds).

• CR-A04-086 (code, Info) — forgelm/cli/ holds 13.4K LOC of business logic although the standard calls the CLI a thin shim; operator identity is resolved in three divergent copies
  severity Info · reachability default · intent intended-but-harmful · confidence likely · single reviewer · not reproduced (code trace)
  Failure scenario. The identity formats drift further. Business logic in cli/ is not covered by the compliance and library test seams.
  Where. forgelm/cli/_pipeline.py:1, forgelm/cli/subcommands/_purge.py:1, forgelm/cli/subcommands/_approve.py:121-186, forgelm/compliance.py:192-232
  Recommended fix. Add the pipeline orchestrator to the ownership table or move it, and share one identity helper.

• CR-A04-087 (code, Info) — Delete and compress checkpoint actions are unreachable from CLI and config
  severity Info · reachability documented-opt-in · intent intended-and-acceptable · confidence likely · single reviewer · reproduced end to end
  Failure scenario. An operator reading architecture.md looks for a way to enable compression in YAML and finds none.
  Where. forgelm/cli/_training.py:152, forgelm/utils.py:57-128, docs/reference/architecture.md:90,175
  Recommended fix. Say 'library API only' in architecture.md or add a config field.

• CR-A04-088 (code, Info) — Two-stage chain through a real ForgeTrainer on tiny-qwen is feasible offline in seconds and exposes the adapter-chain defect
  severity Info · reachability default · intent intended-and-acceptable · confidence likely · single reviewer · reproduced end to end
  Where. tests/test_pipeline_orchestrator.py:95, tests/test_integration.py:1
  Recommended fix. Add an offline integration test that builds a tiny random Qwen2 model and tokenizer in tmp_path and runs a 2-stage sft->dpo pipeline through the real ForgeTrainer, asserting the stage-2 model carries the stage-1 adapter (or merged weights).

• CR-A04-089 (code, Info) — Existing tests pin the very designs two verified findings criticise, so their fixes will break or need those tests rewritten
  severity Info · reachability unreachable · intent intended-and-acceptable · confidence likely · single reviewer · reproduced end to end
  Where. tests/test_pipeline_orchestrator.py:927, tests/test_pipeline_orchestrator.py:1353, forgelm/cli/_pipeline.py:527
  Recommended fix. When fixing M-A04-013, change test :927 to write a real approve event; keep a negative test for reject-then-removed staging.

• CR-A04-090 (code, Info) — Removing lora.use_dora at DEPRECATION_REMOVAL_VERSION turns the dry-run (and training/compliance/model card) into an AttributeError; the removal guard tracks only version strings
  severity Info · reachability unreachable · intent intended-and-acceptable · confidence likely · single reviewer · reproduced (plumbing only)
  Failure scenario. The v1.0.0 removal PR deletes the field but misses one consumer. Every dry-run or training run then crashes with AttributeError. Under the already-reported missing CLI catch-all, that is exit 1 with a traceback and no JSON envelope.
  Where. forgelm/cli/_dry_run.py:53, forgelm/model.py:238, forgelm/model_card.py:222, forgelm/compliance.py:1536
  Recommended fix. Route all consumers through one resolver keyed on lora.method now, so the removal touches only config.py.

• CR-A04-091 (code, Info) — Several kills are single-test or shallow: M3 and M2 rely on one parametrised class; M5/M6/M7/M8a/M9 each have 1-2 killers
  severity Info · reachability unreachable · intent intended-and-acceptable · confidence likely · single reviewer · reproduced end to end
  Failure scenario. Deleting the single killing test, or its parametrisation, would silently remove protection for M6, M7 or M8a.
  Where. tests/test_pipeline_config.py::TestSafetyConfigTrackCategoriesFieldsSet, tests/test_config.py::TestSafetyGateValidation, tests/test_eu_ai_act.py::TestForgeConfigCompliance, forgelm/config.py:1792-1793
  Recommended fix. No action required. The tests are meaningful, and the redundant allow_inf_nan on min_valid_fraction can be left as defence in depth.

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.}