chore: bump acp-kernel to 0.0.34

[ranxianglei]

Bumps acp-kernel 0.0.32 → 0.0.34 (two releases: 0.0.33 responses item.id fix + 0.0.34 inline <think> normalization).

What the proxy gains

• Inline demoted-thinking normalization (kernel PR fix: reap decompress temp files (bounded retention) #112): openaiToCore now splits <think>\n…\n</think> / <thinking>…</thinking> / ```thinking fences at offset 0 of assistant content into a separate reasoning core message when no `reasoning_content` field is present. Providers that demote thinking inline (glm / qwen3 / deepseek / kimi via openai-completions) previously folded those turns as one opaque text blob:
  • reasoning-pair extraction now works for inline turns,
  • fingerprint/identity space is shared with the reasoning_content form, so omp restart replay (issue release v0.1.25 #64 demoted variant) survives without any mirror retry — omp PR feat(mitm): auto-discover model domains from client configs (no hardcoding) #125 drops its omp-side retry accordingly.
• Responses reasoning identity fix (kernel PR 开放尽可能配置的参数 #107, 0.0.33): reasoning core id derived from reasoning text instead of host item id.

Protocol/serialization decisions keep consolidating in the kernel; the proxy stays a thin mitm + adapter shell (this PR is dependency-only).

Tests

• npm test 512/512 pass
• npm run typecheck clean