GPU eval-equivalence verification for #7b (LoRA-only) + #13 (eval subprocess isolation) before merge

Two deferred fixes were implemented + reviewed + CPU-tested but are default-OFF and GPU-unverified. Their opt-in paths must be confirmed to NOT change eval numbers on a real Qwen-7B run before they're merged to main / their flags are flipped on. Both live on wf/fix-13 (which is stacked on wf/fix-7b, so a checkout of wf/fix-13 contains BOTH).

Branches: wf/fix-7b (f5f872f4 + guard fix a6456933), wf/fix-13 (48e191aa). To test: provision a pod, git checkout wf/fix-13.

What to verify

#7b — EPM_MATERIALIZE_MERGED=0 (LoRA-only) eval-equivalence

Train one real condition both ways and compare eval numbers:

• Default (EPM_MATERIALIZE_MERGED unset → merged 7B materialized, current behavior) vs opt-in (=0 → LoRA adapter kept, merged-on-demand / adapter-swap).
• Compare across every eval surface: capability (ARC-C logprob + per-persona ARC/HellaSwag), alignment (Betley emergent-misalignment), generation.
• Known deferred gaps to resolve here: in adapter mode the implementer skipped-with-loud-log the lm-eval OOD vLLM path and the per-persona ARC/HellaSwag merged-only paths, and capability.py adapter mode uses PeftModel.from_pretrained WITHOUT merge_and_unload (claimed numerically equivalent up to fp arithmetic order for the argmax-logprob accuracy metric — confirm empirically). Decide whether those paths need full adapter wiring or the skip is acceptable.
• Acceptance: eval numbers match (exactly, or within a documented tolerance) → safe to merge + consider flipping default. The disk payoff is ~15-30 GB → ~0.1-0.4 GB per phase.

#13 — EPM_ISOLATE_EVAL=1 (subprocess) eval-identity

• Run one real eval both ways: in-process (default) vs isolated (=1, eval runs in a fresh subprocess via run_isolated).
• Confirm reported eval numbers (capability logprob/accuracy, alignment, generation) are byte-identical (or within documented tolerance) across the subprocess boundary on real weights. CPU tests only proved STRUCTURE identity, not float identity.
• Also measure the wall-time cost (the child re-loads model weights per phase — that's the point, clean GPU per phase, but quantify it) and confirm WandB run continuity across parent/child is acceptable (child writes eval JSON to phase_dir + returns the fragment; it does not log to the parent's live WandB run).
• Acceptance: numbers match → safe to merge + flip on for OOM-prone runs. This closes the vLLM-orphan-worker OOM root cause (#399/#397/#396).

Follow-up (separate, smaller)

• cleanup_vllm hardening at the vLLM↔HF boundary (eval/generation.py) — the clearly-safe teardown fix for the orphan-worker class, deliberately left out of #13 (outside that workstream's owned files). Owns generation.py.

Provenance

Both fixes produced by the fix-deferred-risky-3 workflow (2026-05-28); default-OFF so merging the code is low-risk, but the opt-in paths (the actual value) are unvalidated without GPU. #15 from the same batch is already merged (no GPU needed).