PREVIEW8.md

Preview.8 release contract

Invitation-only preview: use v0.1.0-preview.8 only when its non-draft GitHub prerelease provides the Linux/x64 archive and checksum together with the immutable image matrix in RELEASES.md.

Whole-node GPU setup

--all-gpus is an explicit shortcut for a dedicated execution node. It derives the complete locally visible GPU UUID/CDI set, GPU count, and aggregate observed VRAM. Multi-GPU bundles default to SAME_NODE; P2P_REQUIRED remains an explicit stronger topology requirement. Punch never silently advertises all visible GPUs because a shared node may already use some devices elsewhere.

Offer lifecycle UX

An enrolled Provider reuses the same Provider and machine identities for session renewal, serving, draining, setup rollover, and withdrawal. A fresh setup reference creates another immutable offer; it does not require another invitation or re-registration.

withdraw pauses one exact owned offer only while it remains unreserved and LISTED. It preserves audit history. Control serializes withdrawal with Buyer purchase, so whichever transaction reserves or pauses the offer first wins and capacity already reserved for a Buyer cannot be withdrawn. The offer digest is bound transactionally by Control; an optional explicit digest adds a strict stale-version check.

Retained safety and compatibility

• Preview.8 retains Preview.7 setup rollover, advisory locking, GPU-set canonicalization, and the NVIDIA_CUDA_12_8_1_V2 compatibility class.
• CPU-only, single-GPU, manual atomic 2–8 GPU selection, SAME_NODE, and P2P_REQUIRED remain supported.
• Provider claim and relay polling use bounded timeouts. Mutating setup and withdrawal requests remain idempotent and use stable operator references.
• Public Provider errors expose only bounded codes and status metadata, never raw upstream bodies or credentials.

No release claim is made from source alone. Installation is supported only from the published archive after verifying SHA256SUMS.