apr on GPU: refuse by name before OOM, never fall back to CPU silently #5325

Scope: What apr does when the GPU path cannot run: silent CPU fallback, refusing only after a full CPU rerun, OOM instead of a pre-flight refusal, session headroom that doubles context, effective-config misreporting residency, and ladder/receipt fields that hide which backend ran.

Why it matters: A user who asked for GPU must know whether the GPU ran. A silent fallback turns a GPU defect into a slow success that no gate sees.

Done when: Every GPU verb either runs on GPU or exits non-zero with a named refusal before allocating, and the JSON receipt names the backend that ran; tested on the model ladder.

Epic: #3994. Folded under triage rule 7b (2026-10-14): each row below was its own issue, closed as not planned and kept here as a sub-ticket. Reopen an issue when its work starts or its PR opens.

Sub-tickets

• apr run --gpu: a refused GPU attempt redoes the whole generation on CPU, then refuses it (exit 14) — refuse before the redo #4969 apr run --gpu: a refused GPU attempt redoes the whole generation on CPU, then refuses it (exit 14) — refuse before the redo (milestone none)
  • accept: the outcome in the title, with a receipt
• dense GGUF CUDA-init fallback to CPU is silent in text mode #4977 dense GGUF CUDA-init fallback to CPU is silent in text mode (milestone 0.71.1)
  • accept: Done when: a dense GGUF run on a cuda build whose CUDA init fails prints one stderr line naming the CPU fallback, without --verbose, as the Qwen3.5 path already
• Small unified-memory CUDA devices: 16 GiB headroom refuses every model, no refusal before OOM, wrong pull hint #4947 Small unified-memory CUDA devices: 16 GiB headroom refuses every model, no refusal before OOM, wrong pull hint (milestone 0.71.0); branch fix/e3-4947-recut3
  • accept: the outcome in the title, with a receipt
• qwen35 session: #4274 headroom doubles a 32k turn to 65536 positions — 27B OOMs and silently falls back to CPU #4443 qwen35 session: serve: Qwen3.5 chat turn 2 re-prefills from position 0; the session extend never engages via /v1/chat/completions #4274 headroom doubles a 32k turn to 65536 positions — 27B OOMs and silently falls back to CPU (milestone 0.71.0); branch fix/4443-on-main
  • accept: the outcome in the title, with a receipt
• effective-config reports Qwen3.5 CUDA residency as 'cpu residency' and drops the F2 parity verdict (v0.69.3) #4374 effective-config reports Qwen3.5 CUDA residency as 'cpu residency' and drops the F2 parity verdict (v0.69.3) (milestone backlog); branch fix/4374-effective-config-qwen35
  • accept: the outcome in the title, with a receipt
• ladder producer: measure_code should read apr code JSON backend.ran and backend.fell_back like measure_chat #4986 ladder producer: measure_code should read apr code JSON backend.ran and backend.fell_back like measure_chat (milestone 0.71.1); branch la-72/4986-code-backend
  • accept: the outcome in the title, with a receipt
• FALSIFY-GPU-010 prose: an unreported completion does not null the backend when another ran on the CPU #5045 FALSIFY-GPU-010 prose: an unreported completion does not null the backend when another ran on the CPU (milestone 0.71.1); branch w71-e3a/5045-gpu010-prose
  • accept: the outcome in the title, with a receipt
• apr code on Qwen3.5 Q4_K_M (4B, 9B) on CUDA: does it complete a scripted edit-and-verify task on lambda and gx10? UNMEASURED, and its driver runs through apr serve --gpu (#3571) #3719 apr code on Qwen3.5 Q4_K_M (4B, 9B) on CUDA: does it complete a scripted edit-and-verify task on lambda and gx10? UNMEASURED, and its driver runs through apr serve --gpu (apr serve cannot load Qwen3.5 at all — 'Architecture qwen35 … runs through Qwen35Model'; goal A's floor and the P0 decide-lane design both name apr serve #3571) (milestone 0.71.0); branch w71-e3a/3719-code-on-main
  • accept: the outcome in the title, with a receipt