feat(ai): round-trip Anthropic redacted thinking blocks

[rekram1-node]

Problem

The Anthropic Messages protocol silently dropped redacted_thinking content blocks: they fell through every branch of the stream parser, and there was no assistant-lowering path to send them back. Anthropic requires redacted thinking blocks to round-trip verbatim, so multi-turn thinking + tool-use conversations that hit safety filtering would lose reasoning continuity (or get rejected on the follow-up request).

Solution

Model redacted thinking the same way Vercel's @ai-sdk/anthropic does:

• Stream: a redacted_thinking content_block_start emits reasoning-start with providerMetadata: { anthropic: { redactedData } } and empty text; the existing content_block_stop handling closes the part, and response assembly carries the metadata into the final ReasoningPart.
• Lowering: reasoning parts with a signature (part.encrypted or anthropic.signature metadata) still lower to thinking; signature-less parts carrying anthropic.redactedData lower to { type: "redacted_thinking", data } verbatim. Same precedence order as the AI SDK.
• Schemas: new AnthropicRedactedThinkingBlock in the assistant block union; data field on the stream block schema.

Tests

• round-trips redacted thinking as redacted_thinking blocks — mixed redacted + signed reasoning parts lower to the correct block types in order.
• parses redacted thinking into empty reasoning with redactedData metadata — stream fixture asserts reasoning-start metadata and the assembled empty-text reasoning part.

bun test (330 pass) and bun typecheck clean from packages/ai.