GCP lane: FLEX_START + a3-highgpu H100 intents; enable SPOT to use the idle preemptible A100 quota

Provenance

Originating user prompt (2026-06-12, quota research session): "See how you can get better quota for GCP (H100, H200, everything) search deeply" — deep-research report delivered in-session; this task captures the code-side follow-ups (items 4+5 of the action list).

Context

Deep-research findings (verified against Google docs fetched 2026-06-12):

• The 1-4 GPU H100 shapes (a3-highgpu-1g/2g/4g, GA Jan 2025) cannot be created on-demand at all on Compute Engine — Spot VMs or DWS flex-start VMs only. On-demand NVIDIA_H100_GPUS quota is irrelevant for them.
• DWS flex-start consumes PREEMPTIBLE GPU quota, runs uninterrupted up to 7 days, and works as a single-VM create: gcloud compute instances create ... --provisioning-model=FLEX_START --max-run-duration=3d --request-valid-for-duration=2h. This matches the GCP lane's existing ephemeral --max-run-duration instance pattern almost exactly. Source: https://cloud.google.com/blog/products/compute/introducing-flex-start-vms-for-the-compute-engine-instance-api + https://docs.cloud.google.com/compute/docs/instances/about-flex-start-vms
• The project already holds 8× PREEMPTIBLE_NVIDIA_A100_80GB_GPUS in us-central1, entirely unused — backends/gcp.py pins DEFAULT_PROVISIONING_MODEL = "STANDARD"; the SPOT path is plumbed (spec.extra["provisioning_model"]) but gated on idempotency proofs (gcp.py:333-357).
• Preemptible H100/H100-mega quota requests are filed separately by the user (console); this task makes the lane able to consume them the moment they land.

Deliverables

1. FLEX_START provisioning support in backends/gcp.py: extend ProvisioningModel to SPOT | STANDARD | FLEX_START, render --provisioning-model=FLEX_START + --request-valid-for-duration in render_create_argv, verify interaction with --max-run-duration (≤7d cap) and --instance-termination-action=DELETE (flex-start constraints per docs — verify, don't assume). Golden-test the argv.
2. H100 intent mappings: add a3-highgpu-1g (1× H100) / a3-highgpu-2g (2× H100) entries to INTENT_TO_MACHINE (e.g. opt-in intents lora-7b-h100 / eval-h100, or a gpu_kind override) — gated on a quota pre-check (same pattern as the #608 quota pre-check) so a zero-quota project fails loud at selection, not mid-provision.
3. SPOT idempotency proofs + opt-in enablement for A100: validate the preemption→fresh-idempotent-re-run contract the gcp.py docstring describes (artifacts pushed off-VM during the run survive; re-run is idempotent), then allow provisioning_model=SPOT for A100 intents to use the idle 8-GPU preemptible pool. DEFAULT_PROVISIONING_MODEL stays STANDARD.
4. Update the backend-router docs/markers (epm:backend-selected reason field) to name which provisioning model + quota pool a launch used.

Notes

• Item 3 is usable today (quota already granted). Items 1-2 become usable once preemptible H100 quota lands (user console action, tracked outside this task).
• Flex-start is queued/best-effort (request validity 90s-2h) — selector should treat "no capacity within validity window" as a provisioning failure that falls through to the next lane, consistent with the router contract.