keryx-miner-supr v0.6.5.4
Three robustness fixes — no protocol/consensus change, drop-in over v0.6.5.x (mining.subscribe user-agent stays keryx-miner-supr/0.6.5).
Fixes
- Stratum job watchdog — recover automatically when the pool connection wedges. If the pool keeps sending non-job traffic (vardiff/keepalives) but stops delivering block templates, the miner now forces a reconnect after
KERYX_POOL_JOB_TIMEOUT(default 300 s) instead of running "alive" at 0% GPU with zero shares. (Observed live: a rig sat idle ~15 h while reporting a bogus high hashrate.) - Disk pre-check for the possession tree — the PoM tree is ~2× the model size and is built next to it. On a too-small disk the build used to ENOSPC-loop and surface only as endless "Workers stalled". It now fails fast with an actionable message ("needs ~N GB free, only ~M available — free up disk or run a smaller --tier"). UNIX-only check (
df); the Windows build is unchanged. - Rate-limited index-build failure log so the real reason (e.g. disk) is visible without flooding.
Per-tier disk requirement (model + possession tree)
--light ~12 GB · --tier default ~15 GB · --high ~60 GB · --very-high ~85 GB free disk.
Build lines
modern (CUDA 12.9, R575+), legacy (CUDA 12.4, R550+), pascal (sm_60; run --cpu-inference --light). Each in hiveos / smos / mmpos / linux.