fix: update Pi runtime for GPT-6 models

[QAyong]

Why

The desktop app currently pins the Pi runtime at 0.86.1. That runtime does not expose the newer GPT-6 Sol and GPT-6 Luna models through the OpenAI OAuth provider. Pi 0.87.1 adds those model definitions.

Changes

• Update @earendil-works/pi-ai and @earendil-works/pi-coding-agent from 0.86.1 to 0.87.1.
• Refresh pnpm-lock.yaml and carry the existing local patches forward to the new package versions.
• Keep only the patch hunks that are still needed; the cache-warmer changes are already included upstream in 0.87.1.
• Add GPT-6 Astra, Luna, and Sol to the OpenAI Fast/priority service-tier allowlist.

No authentication flow, endpoint, or provider API behavior is changed.

Validation

• vue-tsc -b --pretty false passed as part of the desktop build.
• Windows NSIS packaging passed with pnpm --filter @uselexora/lexora-buddy package:windows.
• The resulting installer includes the 0.87.1 Pi runtime and GPT-6 model definitions.

Artifact produced locally: Lexora-Buddy-0.9.1-windows-x64.exe.

The local host is running Node 24.12.0 while the repository declares Node >=26; engine strictness was disabled only for this local build. The installer is not committed to the repository.

[shanyuhai123]

这部分的直接升级风险比较大，引入新版本新能力会导致很多改动，例如 #148 pi 在那个版本引入了 “缓存预热” 的能力，这对于当前涉及大量附属问题。

如果是为了获取新模型，当前得调整拉取模型的机制，官方 /v1/models 接口更优先，pi 内置的模型列表作为兜底。

[shanyuhai123]

此外，希望后续如果有一些新的想法可以先 issue 我们一起讨论下