feat(proxy): clamp max_tokens, fix Responses content shape, surface reasoning

[GoDiao]

Summary

OpenAI-compatible proxy quality fixes for the Anthropic↔OpenAI translation layer.

Changes

• maxOutputTokens provider config + clamp: Add optional maxOutputTokens to SavedProviderSchema/CreateProviderSchema/UpdateProviderSchema (int, positive, ≤10,000,000). When configured, Anthropic max_tokens is clamped via Math.min before forwarding to OpenAI endpoints. Without a cap, the field is still omitted. Fixes Claude Code sending 128K max_tokens that exceeds provider limits (DeepSeek: 8192, etc.).
• OpenAI Responses content shape: Emit input_text/input_image content parts (the correct Responses shape) instead of reusing Chat-shaped image_url objects. Corrects the flush logic so accumulated content is not dropped when a tool block follows text/image.
• Reasoning streaming (Responses): Handle response.reasoning_summary_text.delta / response.reasoning_text.delta / *.done / reasoning_summary_part.done and the reasoning output item, mapping them to Anthropic thinking content blocks. Includes lazy open for providers that omit the output_item.added event. All reasoning blocks are closed before message_delta per Anthropic spec.
• Cache token mapping: Forward prompt_tokens_details.cached_tokens → cache_read_input_tokens on both Chat and Responses streaming paths (previously only output_tokens was reported).
• Shared extractToolResultContent: Dedupe the tool_result content extraction in both transform paths; non-text blocks (images, nested tool_use, thinking) now produce placeholders like [image omitted: …] instead of being silently dropped.

Files

• src/server/types/provider.ts — add maxOutputTokens to schemas
• src/server/services/providerService.ts — thread maxOutputTokens through getProviderForProxy
• src/server/proxy/handler.ts — pass maxOutputTokens to both transforms
• src/server/proxy/transform/anthropicToOpenaiChat.ts — clamp max_tokens, use shared helper
• src/server/proxy/transform/anthropicToOpenaiResponses.ts — clamp max_output_tokens, correct content part shape, use shared helper
• src/server/proxy/transform/toolArguments.ts — new extractToolResultContent
• src/server/proxy/transform/types.ts — add OpenAIResponsesContentPart
• src/server/proxy/streaming/openaiChatStreamToAnthropic.ts — mapStreamUsage with cached tokens
• src/server/proxy/streaming/openaiResponsesStreamToAnthropic.ts — reasoning block handling + cached tokens
• desktop/src-tauri/Cargo.lock — follow dreamcoder-desktop 0.4.0→0.4.5 version bump
• src/server/__tests__/proxy-streaming.test.ts / proxy-transform.test.ts — coverage for the above

Validation

bun test src/server/__tests__/proxy-streaming.test.ts src/server/__tests__/proxy-transform.test.ts
# 98 pass, 0 fail, 300 expect() calls

git diff --check

Note: providers.test.ts / provider-presets.test.ts have 4 pre-existing failures on clean-master (verified by running tests against the unmodified branch); this PR does not touch the preset data or activation env logic those tests cover.