Keep dependency updates compatible with the OpenAI continuation patch

首先复现计划中的 pnpm up "@tanstack/*" --latest 步骤，并针对当前的依赖补丁运行 tests/openai-responses-tool-continuation.test.ts。检查 issue 中描述的 updater workflow、pnpm 配置和补丁处理方式。当 TanStack AI 系列可以更新且不会导致 continuation 行为回归，并且 workflow 能够执行到 lint 和 build 时，即表示完成。

Evidence

The scheduled Update TanStack Dependencies run failed before lint or build on August 30. The exact pnpm up "@tanstack/*" --latest step exits with ERR_PNPM_UNUSED_PATCH because it moves @tanstack/openai-base away from the patched 0.10.5 version.

I reproduced the same failure from current main at 6584b153. In an isolated worktree, allowing unused patches let the update finish and changed @tanstack/ai-openai from 0.22.0 to 0.22.3, with @tanstack/openai-base moving from patched 0.10.5 to unpatched 0.10.8. The focused tests/openai-responses-tool-continuation.test.ts test then failed because the second request had no previous_response_id. The current patch is still required for sequential server-tool turns with provider-side reasoning.

Exact title, body, workflow, patch-package, and keyword searches found no open issue or PR that owns this failure.

Expected impact

The weekly dependency updater is currently guaranteed to fail whenever it tries the latest TanStack AI family. It never reaches the repository lint or build steps, so unrelated TanStack dependency updates also stop. Simply setting allowUnusedPatches or deleting the patch would make the workflow advance while regressing the Builder tool-continuation contract.

Proposed direction

Keep the focused continuation test as the release gate, then choose one owner for the patched AI family:

1. Preferred long-term direction: ship the continuation behavior upstream, update the AI packages to the first release containing it, prove the focused test passes without the local patch, and delete the patch.
2. Until that release exists: upgrade the pinned pnpm from 11.1.0 to a version that supports update.ignoreDeps and exclude the cohesive TanStack AI package set from the scheduled bulk updater. Upgrade that set separately, carrying the patch forward only after its focused test and the full suite pass.

Alternatives and tradeoffs

• Manually port the 321-line patch to every new openai-base release. This keeps AI packages moving, but each port needs careful review against upstream adapter changes and should not be committed straight to main by a broad scheduled update.
• Let pnpm up ignore unused patches. This unblocks the job but silently removes required runtime behavior, and the isolated focused test already proves the regression.
• Exclude only @tanstack/ai-openai. This may still let related direct or transitive AI packages move out of a tested version set, so excluding and reviewing the cohesive AI family is safer.

Decision needed

@tannerlinsley, should we first upgrade pnpm and hold the TanStack AI family out of the weekly bulk updater, or prioritize landing the continuation behavior upstream and keep the weekly job blocked until a release contains it?