R5 — streaming partial-tag state-machine precedent

[weselben]

Question

Read the existing streaming code in this repo to find prior art for
handling state across SSE chunks. Concretely:

• internal/streaming — what utilities exist for partial-buffer
  reassembly?
• How do existing provider adapters handle finish_reason /
  tool_calls delta-reassembly?

Report: which existing utility or pattern should the think-block
translator's chunk-reassembly use (or whether to write a small new
helper that follows the same idiom). One paragraph; one file path.

[weselben]

Resolution

Existing prior art in this repo:

• internal/streaming/stream_buffer.go — pooled byte FIFO for short-lived stream converters (used by OpenAIResponsesStreamConverter).
• internal/providers/responses_converter.go — per-stream typed state machine consuming chat SSE chunks, the closest analogue.
• internal/anthropicapi/stream.go — block-open/block-close state machine for Anthropic SSE dialect.

Decision: the think-block translator uses a dedicated State (string-builder based) rather than reusing StreamBuffer, because the unit of buffering is semantic tag boundaries, not raw bytes. Implemented in internal/thinkextract/thinkextract.go (PR #41).