fix(opencode): preserve V1 tool attachment filenames

[opencode-agent]

Summary

Fix V1 tool attachment serialization for Responses providers, including Bedrock Mantle.

V1 emits attachments as legacy AI SDK media parts. The SDK converts those to file-data without retaining a filename, so a PDF tool result reaches Responses as filename: "data". Replaying that attachment on later turns keeps sending the same extensionless filename, even when the latest tool only returns text.

• Emit explicit image-data and file-data parts.
• Preserve file attachment names and use attachment.pdf for unnamed PDFs, including existing session history.
• Add request-serialization regressions for Bedrock Mantle and OpenAI Responses: plain text, images, named/unnamed PDFs, and a PDF replayed before a subsequent glob result.

Scoped to the V1 serializer; no V2, provider-routing, dependency, or storage changes.

Validation

• Confirmed the six PDF regression cases fail before the fix and pass afterward; the four text/image controls pass in both cases.
• bun test test/session/message-v2.test.ts test/provider/transform.test.ts test/provider/amazon-bedrock.test.ts test/session/llm-native.test.ts --timeout 30000 — 511 pass.
• bun typecheck — passed in packages/opencode.
• Prettier and git diff --check — passed.
• Local differential SDK serialization checks preserve existing payloads for Anthropic images/PDFs, Bedrock Converse images, Gemini images/PDFs, and xAI images.

Request serialization was tested with intercepted transports, not live provider calls.

Requested by: @rekram1-node (Aiden via Slack)