[Wave 3] Medium — Docs: docs/product_strategy.md #580

Wave 3 · 5 finding(s) in 5 fix package(s). One pass over this document; keep the Turkish mirror in sync.

Findings

• DOC-D2B09-010 (docs, Medium) — docs/product_strategy.md: wrong detail — the final model lands in final_model.staging/ — operators run forgelm approve <run_id> ... or forgelm reject <run_id> to discard
  severity Medium · already tracked: KNOWN-OPEN DOCS-03/DOCS-04 (S15 staging path) — docs/roadmap/phase-16-trust-surface-hardening.md:773
  Where. docs/product_strategy.md:39 (Turkish mirror: docs/product_strategy-tr.md:39) · type: wrong detail
  The document says. the final model lands in final_model.staging/ — operators run forgelm approve <run_id> ... or forgelm reject <run_id> to discard
  What the code does. The staging directory is final_model.staging.<run_id>/. reject does not discard: it keeps the staging directory for forensic review and records human_approval.rejected. Both approve and reject require --output-dir.
  Code references. forgelm/config.py:1222, forgelm/trainer.py:1521, forgelm/cli/_parser.py (reject help: 'left untouched')
  Proposed correction. Rewrite as: "the model lands in final_model.staging.<run_id>/; run forgelm approve <run_id> --output-dir <dir> to promote it, or forgelm reject <run_id> --output-dir <dir> to record a rejection (the staging directory is kept for forensic review)."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D2B09-011 (docs, Medium) — docs/product_strategy.md: stale — Meaningful exit codes: 0 (success), 1 (config/validation error), 2 (training error), 3 (evaluation/safety failure), 4 (awaiting human approval)
  severity Medium · already tracked: KNOWN-OPEN C31-EXIT-CODE-PROSE — docs/roadmap/phase-16-trust-surface-hardening.md:782-785
  Where. docs/product_strategy.md:46 (Turkish mirror: docs/product_strategy-tr.md:46) · type: stale
  The document says. Meaningful exit codes: 0 (success), 1 (config/validation error), 2 (training error), 3 (evaluation/safety failure), 4 (awaiting human approval)
  What the code does. The public contract is 0-6: 5 = wizard cancelled and 6 = integrity failure from verify-. The document's own competitive table (L96) says 0/1/2/3/4/5/6.
  Code references. forgelm/cli/_exit_codes.py:10-39
  Proposed correction. Append: ", 5 (wizard cancelled), 6 (integrity failure — a verify- command compared an artefact and it did not match)."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D2B09-012 (docs, Medium) — docs/product_strategy.md: stale — Meaningful exit codes for pipeline orchestration: 0 (success), 1, 2, 3, 4 (awaiting human approval)
  severity Medium · already tracked: KNOWN-OPEN C31-EXIT-CODE-PROSE — docs/roadmap/phase-16-trust-surface-hardening.md:782-785
  Where. docs/product_strategy.md:46 · type: stale
  The document says. Meaningful exit codes for pipeline orchestration: 0 (success), 1, 2, 3, 4 (awaiting human approval)
  What the code does. The training path emits only 0-4, but the public contract also includes 5 (wizard cancelled) and 6 (integrity failure) (forgelm/cli/_exit_codes.py:21,40,42-51). The same document's table at line 96 says 0/1/2/3/4/5/6.
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D2B09-013 (docs, Medium) — docs/product_strategy.md: stale — Wizard mode (forgelm wizard) ... Quickstart templates (Phase 10.5) will provide working configs
  severity Medium · already tracked: KNOWN-OPEN DOCS-11 docs/roadmap/phase-16-trust-surface-hardening.md:801-803
  Where. docs/product_strategy.md:74 · type: stale
  The document says. Wizard mode (forgelm wizard) ... Quickstart templates (Phase 10.5) will provide working configs
  What the code does. The parser has 19 subcommands and none is 'wizard'; the wizard is the --wizard flag (_parser.py:1329). The quickstart subcommand exists and five templates ship (forgelm/templates/{code-assistant,customer-support,domain-expert,grpo-math,medical-qa-tr}; quickstart.py:68).
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

• DOC-D2B09-015 (docs, Medium) — docs/product_strategy.md: wrong detail — ForgeLM delegates to bitsandbytes/AWQ/GPTQ/HQQ and existing runtimes
  severity Medium
  Where. docs/product_strategy.md:86 (Turkish mirror: docs/product_strategy-tr.md:86) · type: wrong detail
  The document says. ForgeLM delegates to bitsandbytes/AWQ/GPTQ/HQQ and existing runtimes
  What the code does. Quantization in forgelm/ is bitsandbytes (4-bit NF4 via BitsAndBytesConfig) only; the strings awq, gptq and hqq do not occur anywhere in forgelm/. GGUF export goes through llama.cpp's converter.
  Code references. forgelm/model.py:440-450
  Proposed correction. "ForgeLM delegates quantization to bitsandbytes (QLoRA) and GGUF conversion to llama.cpp; it does not implement or integrate AWQ, GPTQ or HQQ."
  Turkish mirror. the Turkish page has the same error; fix both in the same PR

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.}