[Wave 3] Medium — Security-contract fixes: Other (error-path tests and correctness) #645

Wave 3 · 2 finding(s) in 2 fix package(s). Medium-severity findings that touch the security contract (exit codes, secrets, paths, provenance): do these before the rest of wave 3.

Findings

• CR-A06-023 (code, Medium) — Hard-failure paths of the eval gates are untested: judge SSRF block, benchmark init/execution failure, --benchmark-only exit 3, generative-guard load failure
  severity Medium · reachability default · intent contract-violation · confidence confirmed · independently confirmed · reproduced end to end
  Failure scenario. A change makes a blocked judge endpoint, a crashed lm-eval run or a failed guard load report passed=True, or makes a failed benchmark-only run exit 0. CI cannot tell. This is the 'every non-zero exit path has a test' rule (STD-101) not being met.
  Where. forgelm/judge.py:650-656, forgelm/benchmark.py:177-191, forgelm/cli/_no_train_modes.py:123-125, forgelm/safety/_orchestrator.py:251-263
  Recommended fix. Add run-level tests: judge with a private api_base returns passed=False with the policy reason, lm_eval stubs that raise in HFLM and simple_evaluate, a failing benchmark-only SystemExit(3), and a generative-guard load failure through run_safety_evaluation.

• CR-A10-027 (code, Medium) — Approve on a cross-device output dir can leave a partial final_model/ with no audit event and blocks the retry
  severity Medium · reachability documented-opt-in · intent contract-violation · confidence confirmed · independently confirmed · reproduced with stubs
  Failure scenario. staging on a different mount from final_model (documented EXDEV case), disk fills mid-copy: exit 2, final_model/ contains a subset of the weight shards, staging intact, audit chain says nothing. A downstream export/deploy/chat that defaults to final_model/ would load a truncated model.
  Where. forgelm/cli/subcommands/_approve.py:227-255 (_atomic_rename_or_move; shutil.move fallback at :254), forgelm/cli/subcommands/_approve.py:573-581 (comment promising the gate 'either succeeds with both promotion and audit, or fails with neither'), forgelm/cli/subcommands/_approve.py:565-570 (os.path.lexists(final_path) guard that blocks the retry)
  Recommended fix. On failure in the move branch, remove the partial destination (and record a failed-promotion audit event), or copy to final_model.partial.<run_id> and rename into place.

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.}