fix: preserve Ollama tool calls and pasted approval feedback

[AetherAI3]

Summary

• Give every Ollama tool call a unique ID across model replies and carry that ID through the assistant turn, host event, and tool result.
• Buffer incomplete terminal escape sequences while collecting denial feedback, so split bracketed-paste markers cannot submit a pasted newline or leak text to the next input owner.

Verification

• npm run build
• node --test --test-isolation=none dist/test/brain_ollama.test.js dist/test/approval_feedback.test.js dist/test/bridge.test.js (50 passed)
• git diff --check

Follow-up to the review on #334 by @ability-spec.

[ability-spec]

=)