apr finetune merge: a PEFT adapter, a partial match or a missing lora_alpha gives a wrong merge with exit 0 #4674

apr finetune merge can write a wrong merged model and still exit 0. verify_merged_runnable (line 454) checks only that the output is valid APR v2, has sound metadata and runs, so it passes each of these cases. The 0.72 look-ahead desk found this by reading the code; it has not been run yet (C277: no runs before 0.70.1 is live).

The code is the same at main 316dee2 and at the 0.70.1 release commit 04332f8 (crates/apr-cli/src/commands/finetune_display_next_validate.rs, line numbers at 04332f8):

1. A PEFT-named adapter matches nothing, and the output is the unchanged base. PEFT names its tensors …lora_A.weight / …lora_B.weight. The candidate names (lines 205-230) are only {base}.lora_a and lora.{layer}.{proj}.lora_a, and the zero-match refusal counts only names ending in .lora_a (line 590). A PEFT adapter therefore counts 0, is not refused, and gives the base back with exit 0.
2. Partial matches are written. gguf_proj_to_hf (line 192) maps only the 7 llama projections. Any other projection is skipped without a word, so the model is merged on some layers only.
3. A missing lora_alpha defaults to 16.0 (lines 88 and 101, in read_adapter_lora_params). That silently changes the merge scale (alpha/rank) for any adapter trained with another alpha.

Falsifier text: QQE-009 on branch la-72/r15-receipt-ext @ed73bafff8.

Done when:

• each of the three cases either refuses with a non-zero exit or merges correctly;
• verify_merged_runnable fails on a merge that changed no weight;
• a test covers each case and turns RED when the fix is reverted.

Not in Qwen3.5 only: this is the general merge path. No work before 0.70.1 is live (C277.3).

Reported by la-0.72 (inbox 20:17Z). The lines above were checked against 04332f8 by the cop.

Agent: aprender-27