[Fix] Gemini judge fails after tool calls

[roomote-roomote]

​Opened on behalf of @mrubens. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Consolidate split OpenRouter reasoning details before OpenCode replays Gemini 3 assistant tool-call turns.
• Preserve genuine opaque encrypted blocks and their tool-call IDs unchanged.
• Add Google's documented validator-bypass sentinel only when an older or partial tool-call history has no encrypted reasoning block.
• Cover preservation, deduplication, idempotent fallback injection, and provider/model scoping in the shared plugin used by both standard tasks and prompt-only subagent runtimes.

Why this change was made

Gemini judge subagents could fail after a tool call with HTTP 400 when OpenRouter could not find replayable reasoning details on the reconstructed assistant message. OpenCode stores provider metadata per part while the OpenRouter serializer reads the first reasoning-details array it encounters, so partial or split histories could omit the encrypted block required for continuation.

Impact

Judge and other Gemini 3 subagent tool loops can continue across persisted OpenRouter turns without exposing private reasoning or credentials. Existing signed metadata remains authoritative; the fallback is limited to otherwise unsigned OpenRouter Gemini 3 tool-call turns.

Focused tests, package type checks, lint, formatting, and the full pre-push lint/type/Knip gates passed. An isolated OpenCode 1.18.30 replay verified both opaque metadata preservation and unsigned-history fallback on openrouter/google/gemini-3.8-flash; this was a local mocked-provider verification, not a live OpenRouter request. The unrelated existing OpenCode plugin-seed test remains failing on develop and in isolated execution.

[roomote-community]

No code issues found. See task

_{Reviewed 10d5803}