fix: correct gemini interactions streaming

[ADabbler]

Summary

Fix Gemini Interactions streaming handling and interrupted-generation recovery.

• Parse the official event_type field from Gemini Interactions SSE events.
• Route step.delta events with delta.type: "thought_summary" to the reasoning stream rather than normal assistant text.
• Reconstruct function-call arguments from step.start and subsequent arguments_delta events.
• Read official interaction usage fields, including thought-token usage.
• Remove the failed-run marker that could prevent a server-pending generation from resuming after the initial browser stream disconnected.

Change type

• Bug fix

Affected areas

• Frontend / UI
• Backend / API
• Conversations / streaming
• Model routing / providers

Verification

• go test ./internal/infra/llm -count=1
• pnpm check
• pnpm test
• pnpm build
• git diff --check

Screenshots, API examples, or logs

The updated stream test fixture covers the official event shape:

{
  "event_type": "step.delta",
  "delta": {
    "type": "thought_summary",
    "content": {
      "type": "text",
      "text": "..."
    }
  }
}

No UI layout changes.

Configuration, migration, and compatibility notes

• No database migration, deployment configuration, environment variable, or HTTP API contract change.
• No generated artifacts are required.
• Existing model settings are unchanged.

Documentation

• Documentation is not needed for this change.

Security and privacy

• No secrets, tokens, credentials, local config, or personal data are included.
• User data access remains scoped by authenticated user context unless an admin-only path explicitly requires broader access.
• Security-sensitive behavior was reviewed, including authentication, authorization, provider routing, file processing, billing, and admin APIs where relevant.

Checklist

• I searched existing issues and pull requests.
• Changes are focused and do not include unrelated refactors.
• Tests or static verification were run where practical.
• User-facing behavior, deployment steps, API contracts, or configuration changes are documented.
• Generated artifacts are included only when this project explicitly requires them.
• Caches, build output, .pyc files, .env files, and local storage data are not committed.

[chenyme]

Additional fixes

• Preserve text, thought summaries, and thought signatures already present in step.start events.
• Keep function-call reconstruction state local to the Gemini Interactions stream instead of storing protocol-specific state on the shared output type.
• Preserve the observed service_tier across later usage events that omit it.
• Handle official streamed image and video deltas so generated media is retained.
• Parse Google Search, Code Execution, and URL Context call/result steps.
• Merge native tool calls and results by call_id, retain citations, and count each server-side tool invocation only once.
• Notify the frontend when the local active-run set changes so a disconnected server-pending generation can resume without waiting for an unrelated render.

Additional verification

• Compared the implementation against the current Gemini Interactions API reference and streaming documentation.
• Sent an unauthenticated request to the live Google Interactions endpoint and confirmed the official endpoint and API-key authentication behavior through its expected HTTP 403 response.
• Added regression coverage for:
  • step.start text and thought content
  • thought signatures
  • streamed function arguments
  • cumulative usage and cached tokens
  • service-tier preservation
  • native server-side tools
  • streamed images and videos
  • citation extraction and non-duplicated tool usage
• go test ./internal/infra/llm -run 'GeminiInteraction|GeminiInteractions' -count=1
• go test -race ./internal/infra/llm -run 'GeminiInteraction|GeminiInteractions' -count=1
• pnpm check
• pnpm test
• pnpm --filter @deeix/web build
• pnpm --filter @deeix/api build
• git diff --check

Compatibility notes

• No database migration, HTTP API contract, environment variable, or deployment configuration change is required.
• Existing model settings remain unchanged.
• The Gemini Interactions API is still in beta, so future upstream schema changes may require further adapter updates.
• An authenticated live generation was not run because no Gemini API credential was available in the verification environment.