fix: normalize ToolCall arguments to a map when serializing tool calls

[brainlid]

Closes #580

Problem

When an assistant Message reaches a chat model with a ToolCall whose arguments is still a string (rather than a parsed map), the serialized request sends the tool input as a string and the provider rejects the whole request. On Anthropic:

messages.N.content.M.tool_use.input: Input should be an object

ToolCall.arguments is only parsed from string → map by validate_and_parse_arguments/1, which runs only when status: :complete (via ToolCall.complete/1). Any flow that puts a ToolCall into an outgoing message without going through that completion path still holds a string:

• a persisted/streamed tool call rehydrated as status: :incomplete (agent frameworks that serialize state and replay history)
• an interrupted stream where the tool_use JSON was truncated before content_block_stop
• a zero-argument tool call whose arguments accumulated as "" rather than "{}"

Anthropic's for_api/1 guarded this with call.arguments || %{}, but that only substitutes for nil/false. "" || %{} is "", and a non-empty JSON string is truthy, so the string was sent verbatim.

Root cause is broader than Anthropic

ChatGoogleAI and ChatVertexAI have the same passthrough ("args" => call.arguments) and the same latent bug. (The OpenAI-family providers use Jason.encode!/1, a different code path, and are left unchanged.)

Fix

Add a shared, documented ToolCall.arguments_as_map/1 helper that normalizes any shape into a map:

• a map is returned unchanged
• a JSON-object string is decoded
• nil, empty string, and undecodable / non-object strings become %{}

and use it in the three providers that pass arguments straight through:

• ChatAnthropic — tool_use.input
• ChatGoogleAI / ChatVertexAI — functionCall.args (both the plain and thoughtSignature clauses)

Google and Vertex previously sent args: nil for zero-argument calls; they now send an empty object, which is the correct representation and consistent with the Anthropic behavior.

Backward compatible: maps pass through unchanged, so existing serialization is unaffected.

Tests

• ToolCall.arguments_as_map/1 unit tests + doctests covering map / JSON string / nil / empty / invalid / non-object.
• ChatAnthropic.for_api/1 regression tests for a string-arguments call (→ object) and an empty-string call (→ empty object).
• Full suite green: 35 doctests, 2017 tests, 0 failures.

🤖 Generated with Claude Code