docs(bedrock): post_call guardrails scan /converse-stream pass-through streams as they arrive

docs/pass_through/bedrock.md

@@ -184,6 +184,15 @@ curl -X POST 'http://0.0.0.0:4000/bedrock/model/my-claude-model/converse-stream'
 }'
 ```
 
+#### post_call guardrails on `/converse-stream`
+
+A `post_call` guardrail on the caller's key scans `/converse-stream` output while it streams, the same way it scans `/v1/chat/completions` streams:
+
+- With `streaming_buffer_until_moderated: false` (the default for most guardrails) the proxy forwards frames as Bedrock sends them and scans the text every `streaming_sampling_rate` frames (default 5) and at the end. Text reaches the client before the guardrail's verdict, and masks are not applied. Each scan pauses the stream until the guardrail answers, so a slow guardrail stretches a long answer (a 700-word answer took about 47 s instead of 12 s with the Bedrock guardrail); add `streaming_end_of_stream_only: true` to scan once at the end instead.
+- With `streaming_buffer_until_moderated: true` (the default for the [Bedrock guardrail](../proxy/guardrails/bedrock#streaming)) the proxy holds the stream until it is scanned, then sends it with masks applied per content block.
+- A block after the first frame ends the stream with an AWS `validationException` event, which the AWS SDK raises as `EventStreamError` with the guardrail's message. A block before the first frame returns an HTTP error.
+- Guardrails that only implement `async_post_call_success_hook`, guardrails with `mask_response_content: true`, and post_call policy pipelines still hold the whole stream and scan it once.
+
 ### Supported Bedrock Endpoints with config.yaml
 
 When using models from config.yaml, you can call any Bedrock endpoint:

docs/proxy/guardrails/bedrock.md

@@ -202,7 +202,7 @@ Chunks now stream to the client as they arrive, and one OUTPUT scan runs over th
 
 With `streaming_buffer_until_moderated: false` alone, the guardrail scans the accumulated response every `streaming_sampling_rate` chunks while streaming. Each sampled scan is a separate ApplyGuardrail call over all text so far, so it adds mid-stream latency and repeated Bedrock text-unit charges. Pair it with `streaming_end_of_stream_only: true` unless you need mid-stream blocking.
 
-These settings apply to both `/v1/chat/completions` and native `/v1/messages` streams.
+These settings apply to `/v1/chat/completions`, native `/v1/messages` streams, and the Bedrock `/bedrock/model/<model>/converse-stream` pass-through, where a mid-stream block ends the stream with an AWS `validationException` event that the AWS SDK raises with the guardrail's message.
 
 ## Contextual Grounding