Release v8.0.0

Release v7.0.1

Release v7.0.0

Add candle adapter for neural inference on streams (#239)

Implements the candle adapter scoped in issue #239: a transform-style
adapter that runs a candle model over a wingfoil stream via a single
`candle_infer` factory.

- The model is supplied as a `forward` closure (candle has no uniform
  load-from-path type), keeping the adapter architecture-agnostic.
- Two execution modes via `CandleMode`: `Inline` (forward runs in cycle(),
  deterministic, both run modes) and `OffThread` (worker thread + ReceiverStream,
  real-time only) for heavy models.
- CPU backend only in v1; GPU features deferred so `--all-features` CI does
  not require CUDA/Metal toolchains (see issue #239 §9).
- Reuses the ReceiverStream primitive for off-thread wakeup; its cfg gates
  (nodes/mod.rs, nodes/channel.rs) are extended to include the candle feature.
- Example, READMEs, and adapter CLAUDE.md added. Deviations from /new-adapter
  (no read/write split, no integration tests/Python bindings) documented.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01L84sh4wVgVMuLCjmtPSki8

Release v6.0.5

Release v6.0.4

Release v6.0.3

Release v6.0.2

Release v6.0.1

Release v6.0.0