fix(prism-lium): abort rent when rent_gpu_count > 1

[echobt]

Summary

Hard-abort Lium POST /rent when rent_gpu_count exceeds the requested width (InstanceSpec.gpu_count / Proof HarvestLimits.gpu_count, default 1).

Prod 2026-09-08 rented an 8× NVIDIA B200 (rent_gpu_count=8) then hit CREATION_FAILED. The previous abort only covered 8× 5090. NCU / whole-host B200 rent_count(1)→8 was still posted.

• Skip (do not POST) any candidate with rent_gpu_count > spec.gpu_count, then continue so a later split 8× B200 can still rent 1×.
• Solo oversize inventory still errors with the abort message; no HTTP rent body with gpu_count=8.
• Idle split 8× B200 with available_gpu_count still rents 1×.
• provision_refuses_8x_b200_when_requesting_one proves the 8× NCU path is refused.
• provision_skips_ncu_upsell_and_rents_split_1x covers Greptile P1 (2× NCU + split 8× → rent 1×).

Did not invent digests, touch FORCE_SIM/StubWin, or change prod pins.

Greptile

Every PR is reviewed by Greptile before merge. Config: .greptile/.

• Greptile has reviewed this PR; findings are fixed or answered
• If the bot was silent, I commented @greptileai review

Test plan

• cargo test -p prism-lium -p prism-lium-types
• cargo fmt --all -- --check
• cargo clippy -p prism-lium -p prism-lium-types --all-targets -- -D warnings

Risk

Proof harvest will not POST NCU / whole-host upsells that exceed the requested GPU count. A later split-capable offer in the same candidate set can still rent 1×. No deploy pin or emission change.

Naming

I did not rename BASE_* environment variables, deployed host paths
(/opt/base, /run/base, …), GHCR baseintelligence/base package names, or
base-*-v1 cryptographic domain tags, unless this PR’s purpose is a coordinated
cutover documented in docs/NAMING.md.

  

[greptile-apps]

Greptile Summary

This PR generalizes the LIUM pre-rent guard so provisioning refuses any whole-host GPU count greater than the requested width and adds coverage for 2× and 8× NCU B200 offers.

• Prevents oversized B200 and 5090 rent requests from reaching POST /rent.
• Preserves one-GPU rental from explicitly split-capable B200 hosts.
• Updates offer-width documentation and provisioning tests.
• The new hard return can prematurely reject a candidate set containing a later valid split offer and bypasses no-capacity recovery.

Confidence Score: 4/5

The PR is not safe to merge until an oversized candidate is handled without skipping later valid offers and exhausted candidates retain no-capacity recovery semantics.

The rent guard successfully prevents oversized POST bodies, but its immediate generic error can turn an otherwise satisfiable or recoverable capacity condition into a permanent provisioning failure.

Files Needing Attention: crates/prism-lium/src/client.rs

Important Files Changed

Filename	Overview
crates/prism-lium/src/client.rs	Generalizes the pre-POST over-width guard and adds B200 refusal tests, but aborts the entire candidate loop on the first oversized offer.
crates/prism-lium-types/src/types.rs	Updates documentation to explain that whole-host NCU rent counts can exceed the requested width and are refused by the client.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[List and sort matching offers] --> B[Examine next candidate]
  B --> C{rent_count > requested?}
  C -->|Yes, current implementation| D[Return generic API error]
  D --> E[Later valid offers skipped]
  D --> F[Not classified as no-capacity]
  C -->|No| G[POST requested rent]
  G --> H[Wait for running pod]

Loading

_{Reviews (1): Last reviewed commit: "fix(prism-lium): abort rent when rent_gp..." | Re-trigger Greptile}