git:20260928.eb5d1db to git:20261008.aafd09c

9 added, 2 removed. Audit A to A.

---

name: gpu-allocation-fallback

description: Track typed GPU placement failures in the NPA agent and offer a consent-gated on-demand-to-preemptible fallback after repeated failures or deterministic preflight. Use when quota, capacity, Unschedulable GPU, or compatible-product placement blocks an allocation.

---

# GPU Allocation Fallback

Send typed results to `POST /api/agent/gpu-allocation/attempt` with a stable

`logical_allocation`, the requested compatibility/execution/disk invariants,

typed failure evidence, and the identical compatible preemptible candidate.

The grounded route uses zero model tokens.

The optional `success` field must be a literal JSON boolean. Omitting it records

the attempt as unsuccessful, the same as `"success": false`; only

`"success": true` resets the tracked allocation as succeeded. Strings, numbers,

`null`, arrays, and objects are rejected with HTTP 400 without changing fallback

or confirmation-token state.

Only quota/capacity exhaustion, insufficient GPU or Unschedulable, and no

- compatible product/affinity count. Auth, RBAC, network, image-pull, checkpoint,

- application, runtime, cancellation, and timeout failures never count. The

+ compatible product/affinity count. For a Nebius Compute failure, send the

+ operation's `status.details[].code` as the failure code and the violation text

+ as the message: `NotEnoughResources` counts as capacity exhaustion, and

+ `QuotaFailure` counts only when the message names a `compute.instance.gpu.*`

+ quota. Auth, RBAC, network, image-pull, checkpoint, application, runtime,

+ cancellation, and timeout failures never count, even when a placement code is

+ also present. A timeout means the `timeout` code or a message that says the

+ operation timed out; the `NotEnoughResources` message's "VM schedule timeout"

+ wording is a capacity failure. The

default prompt occurs on the third qualifying failure, or immediately when

deterministic preflight proves on-demand cannot succeed and compatible

preemptible capacity is available.

Never switch automatically. Present the returned question and proposed action.

Accept through `POST /api/agent/gpu-allocation/consent` with its single-use,

action-digest-bound `confirm_token`; decline with `accept: false`. A decline

keeps on-demand and suppresses the same evidence until materially new evidence

arrives.

Preserve GPU family/product/count, image/digest, SM and RT-core requirements,

backend, model, workload tier, execution mode, and boot-disk count/bytes. The

state record contains only redacted digests, classification, selected pool, and

consent outcome. Success or a changed logical allocation resets attempt state.

Verify changes with:

```bash

npa/.venv/bin/python -m pytest \

npa/tests/cli/test_agent_gpu_allocation_fallback.py \

npa/tests/cli/test_agent_backend_render.py -q

```