feat: misina/stream — tool-call accumulator + stream.collect() helpers

[productdevbook]

Problem

LLM streaming protocols (OpenAI, Anthropic) deliver tool-call arguments as deltas:

• OpenAI: choices[].delta.tool_calls[] with index, function.name (first chunk), function.arguments (concat across chunks)
• Anthropic: content_block_start → content_block_delta (input_json_delta.partial_json chunks) → content_block_stop

Every consumer rewrites the merge logic. Misina can ship two helpers and a generic accumulator.

Proposal

misina/stream adds:

// Generic stream consumer with reducer
const final = await stream.collect(reducer, initial)

// Provider-specific
import { accumulateOpenAIToolCalls, accumulateAnthropicMessage } from 'misina/stream'

const toolCalls = await accumulateOpenAIToolCalls(sseStream(response))
const message = await accumulateAnthropicMessage(sseStream(response))

Both helpers stop at the appropriate sentinel ([DONE] for OpenAI, message_stop for Anthropic). They're 50-80 lines each.

Acceptance criteria

• Generic collect(reducer, initial) async iterator helper
• OpenAI tool-call accumulator (handles index, name, partial JSON arguments)
• Anthropic message accumulator (content blocks, tool_use input_json_delta)
• Tests with realistic streaming fixtures
• Doc with both providers

Refs

• OpenAI streaming format
• https://platform.claude.com/docs/en/api/messages-streaming

[productdevbook]

Shipped in d68e0fb. misina/stream now exports collect, accumulateOpenAIToolCalls, accumulateAnthropicMessage.