Three fixes:
1. cuCtxSetLimit(STACK_SIZE, 4096) in GpuCuriosityTrainer::new() — the
fused kernel needs ~3KB/thread (6 arrays of 42-128 floats), default
1024B causes stack overflow → async crash → stream deadlock
2. ml crate default features restored to ["minimal-inference", "cuda"]
— was ["minimal-inference"] only, causing #[cfg(not(feature="cuda"))]
gates to fire and block GPU code paths
3. Removed candle-core, candle-nn, candle-optimisers from ml/Cargo.toml
dependencies — Candle was eliminated from source but deps remained.
NOTE: 322 candle references remain in ml/src/ — next commit migrates them.
4. Re-enabled curiosity in smoke test (curiosity_weight back to default)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
c65d6228f4