[Wave 3] Medium — Module sweep: Model loading, dataset preparation & VRAM fit-check (part 2/2) #560

Wave 3 · 1 finding(s) in 1 fix package(s). Fix these whenever a pull request touches this module ("clean the module you touch").

Findings

• CR-A03-049 (code, Medium) — Inline forgelm ... spans in docs are never validated; shipped inline command errors include forgelm export ... --merge, forgelm compliance-only and forgelm wizard
  severity Medium · reachability default · intent contract-violation · confidence confirmed · independently confirmed · reproduced end to end
  Failure scenario. An operator follows the LoRA user manual's tip to merge an adapter for deployment and runs forgelm export ./checkpoints/run --merge. It fails with an argparse usage error, exit 2, and the page never mentions training.merge_adapters or --adapter. An auditor or deployer reading the compliance_summary or iso_soc2 evidence tables tries to reproduce the 'canonical' forgelm compliance-only case and gets 'invalid choice'. CI stays green because no guard reads inline spans.
  Where. docs/usermanuals/en/training/lora.md:140, docs/usermanuals/tr/training/lora.md:138, docs/reference/compliance_summary.md:120, docs/reference/compliance_summary.md:160
  Recommended fix. Fix the docs. In lora.md (EN+TR), replace the command with training.merge_adapters: true, or with forgelm export <base> --adapter ./checkpoints/run/final_model --output model.gguf for GGUF. Replace forgelm compliance-only with forgelm --config <cfg> --compliance-export <dir> in the 10 doc spans and the 5 docstrings. Replace forgelm wizard with forgelm --wizard. Then extend check_cli_h…

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.}