v0.5.0 — Architecture Hardening

Highlights

Safety: 47 lock-poisoning sites replaced with parking_lot. Panics no longer permanently break the gateway process.

Model-aware context: Tape trimming now uses the model's actual context window instead of hardcoded limits. Tool loop stops before overflowing.

Streaming cancellation: /stop actually stops mid-stream via CancellationToken + tokio::select!.

Runtime providers: Register custom LLM providers without forking nexil.

OAuth auto-refresh: 401 responses trigger automatic token refresh with single-flight guard.

See CHANGELOG.md for full details.

Breaking Changes

• HookError::Plugin.hook_point is now HookPoint enum (was &'static str)
• apply_context_budget() accepts context_window: Option<usize> parameter
• nexil bumped to v0.8.0