Preserve canonical metadata, citations, and stream responses

[raphael]

Summary

• preserve provider-authored message metadata with lossless canonical JSON codecs
• replay Bedrock citations natively across character, chunk, and page locations
• make Streamer.Response() the sole terminal stream state and remove the open-ended metadata side channel
• reject Bedrock ToolChoiceNone when tools are defined instead of silently treating suppression as auto-selection

Why

A completed model turn must survive transport, tracing, and replay without reconstruction. Generic metadata transport can coerce numbers, cited Bedrock answers previously failed on the next turn, and a second stream metadata channel allowed terminal usage and citations to diverge from the canonical response. Bedrock also has no provider-level “none” mode when a tool catalog is present, so accepting that request created a false guarantee and allowed forbidden terminal calls.

How

model.MarshalMetadata and model.UnmarshalMetadata preserve JSON numbers and nested provider data exactly. Bedrock rebuilds validated native citation blocks and rejects unsupported forms explicitly. Streaming providers publish content, usage, and stop reason once through Response() after clean EOF; UsageChunk remains the progressive accounting event while a stream is active. Tool suppression now succeeds only for requests with no advertised tools and otherwise fails fast with a precise capability error.

Test plan

• focused model, provider, runtime, and gateway tests
• make test
• make lint
• git diff --check