fix(responses): preserve tool images and bound inline media

[buluw]

Summary

Preserve images returned by Responses tool outputs as multimodal model input while bounding inline image memory and removing historical inline media before replay or session caching.

Closes #3354.
Closes #3355.

Depends on #3362. This is a stacked PR; its new behavior is contained in one independent commit and the displayed diff will narrow after the prerequisite PR merges.

Changes

• Parse text and image parts from string, array, typed-object, and { "content": ... } tool outputs.
• Keep text-only tool output as a string, but emit mixed text/media as OpenAI content blocks so the existing mapper produces Gemini inlineData parts.
• Apply the same tool-output conversion to HTTP Responses and the WebSocket compatibility path.
• Limit Responses inputs to 16 data-URL images, 20 MiB decoded per image, and 32 MiB decoded in aggregate.
• Measure decoded base64 size without allocating a second decoded image buffer.
• Remove inline image/audio data from cached request history, completed output, debug snapshots, and stored HTTP session input while retaining small placeholders.
• Preserve media in the latest user turn, but replace media before that turn before enforcing the current request limits.
• Move owned JSON values through conversion and history paths to avoid repeatedly cloning large base64 payloads.

Files

• src-tauri/src/proxy/handlers/openai.rs
• src-tauri/src/proxy/http_session_store.rs
• src-tauri/src/proxy/mappers/openai/request.rs
• src-tauri/src/proxy/adapters/apply_patch_preflight.rs

Tests

• cargo test --manifest-path src-tauri/Cargo.toml --offline --lib proxy::handlers::openai::stream_peek_tests
• Result: 19 passed, 0 failed.
• Coverage includes tool-output images reaching Gemini inlineData, text-only compatibility, pure-image history placeholders, case-insensitive data-URL metadata, per-image and aggregate byte limits, 16-image limits, and pruning historical images before validating the latest user turn.
• git diff --check passed.

Compatibility

• Text-only Responses tool outputs retain their existing string representation.
• Current-turn images remain available to the model; only older inline media is replaced by a small textual placeholder.
• The hard byte and count limits apply to inline data: image URLs. Remote image URL behavior is unchanged and is not claimed as directly tested here.
• Existing HTTP session storage remains compatible; the session graph and response-ID semantics are intentionally left to a follow-up PR.
• This PR does not change image generation account scheduling, retry policy, or HTTP disconnect handling.