v0.9.6 — escrow modernization, per-GPU fault recovery, download hardening
- Escrow (solo mining): batched claims (up to 87 outputs/TX, fee amortized), claim-response matching by txid, rejected-batch bisection to isolate dead inputs, boot-time state validation against the node (orphaned/reorged ghost entries purged automatically), retry backoff. Ported from upstream v0.4.2/v0.4.3 and live-validated: 95+ solo blocks, concurrent claims accepted, 45 ghost entries purged on a restart test.
- Per-GPU transient-fault recovery: an ILLEGAL_ADDRESS-class CUDA fault on one GPU now drops only that GPU's state (walk ctx + llama engine if hosted there) and rebuilds next cycle — other GPUs keep mining. Previously such errors were silently swallowed and the GPU wedged. OOM and bus-drop errors keep their existing behavior.
- Model download hardening (since BUILD 44): never truncates a multi-GB partial when an IPFS gateway ignores Range; complete-by-tensor-coverage GGUF adoption and stale-.ok repair.
- OPoI text: repetition penalty (256-token window, 1.10) — stops sentence-loop degeneration on small quantized models. Not consensus-relevant.
NVIDIA Linux assets: modern (sm_75+, driver 575+), legacy (sm_70+), pascal (sm_60/61) × HiveOS/SMOS/mmpOS/linux + SHA256SUMS. Windows/macOS follow via CI; AMD assets follow from the AMD line.
⚠️ Known issue (solo mining): keryxd >= 1.4.4 REQUIRED
keryxd v1.4.4 shipped a consensus fix (coin-age maturation on reorg). Pre-1.4.4 nodes have since
DIVERGED from the network ("no chain block in common") and will reject solo blocks with
"Block was not submitted: block invalid" — especially with this release's escrow claims in the
templates. Upgrade your node to v1.4.4+. The Linux binaries (BUILD 46) now detect an old node,
HOLD escrow claims and print an explicit upgrade warning instead of mining into rejections.