Complete local runtime strategy and benchmarking pass

[krahd]

Goal for the August 2026 local-runtime pass:

• Add explicit portable, mac-performance, and auto local runtime profiles without changing general Client(provider="auto") behaviour.
• Include Ollama, BaseRT, vllm-mlx, and oMLX in local runtime discovery/selection; keep raw MLX-LM as a benchmark reference rather than an automatic provider.
• Add conservative runtime capability metadata for streaming, prefix/prompt caching, cancellation, structured output, tool calls, and model discovery.
• Add a conversational local benchmark covering cold TTFT, warm-prefix TTFT, first useful phrase latency, estimated decode rate, context growth, cancellation/stream-close behaviour, and optional process RSS.
• Make local-only construction strict so runtime failures cannot silently become deterministic fallback output.
• Bind an unspecified requested model to a model actually reported by the selected local runtime.
• Reconcile README/API/local-runtime documentation and repository-wide STATUS state.
• Obtain a completely green CI run for PR Add explicit local runtime profiles #23.
• Resolve all current PR review findings.
• Merge PR Add explicit local runtime profiles #23 to main.
• Verify post-merge CI and update tom-work-admin.

Hardware-specific performance ranking remains empirical: the benchmark harness is part of this goal; results for Tomas's M1 Max 64 GB must be recorded only when the benchmark is actually run on that machine, not inferred from CI or external benchmarks.