fix: keep Claude tool results adjacent

[KooshaPari]

Fixes #6026.

Problem:

• Antigravity AgentBridge can send an OpenAI-compatible sequence where a role:"tool" result is separated from the assistant tool call by an intervening user turn.
• The existing OpenAI -> Claude translator only checked that the tool_use id existed somewhere in the conversation. Anthropic requires each tool_result user turn to immediately follow the assistant turn containing the matching tool_use.

Change:

• Reattach valid tool_result blocks immediately after the matching assistant tool_use message.
• Preserve intervening user text after the repaired tool-result turn.
• Drop true orphan tool_result blocks that cannot be matched safely.

Validation:

• npm exec --yes tsx -- --test tests/unit/translator-openai-to-claude.test.ts
• node scripts/check/check-file-size.mjs
• git diff --check

Note: tests/unit/translator-antigravity-to-openai.test.ts hung with no output when run standalone in this worktree, so I interrupted it and did not count it as a pass.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[diegosouzapw]

Thanks @KooshaPari. Two things before this can land: (1) Scope/labeling — it says "Fixes #6026", but #6026 is an Antigravity-native (Anthropic-format) payload that flows through the executor (AntigravityExecutor), which does not import openai-to-claude — so this translator change doesn't actually intercept the reported #6026 case. That one is fixed at the executor layer by #6055. Please drop the "Fixes #6026" (so it doesn't close an issue it doesn't cover) and reposition this as a general openai→claude tool-result-adjacency improvement. (2) Possible new failure — enforceToolResultAdjacency() can emit two consecutive user turns (e.g. [user, assistant(tool_use), user{tool_result}, user "…"]), which Anthropic rejects with roles must alternate. The new test only checks index order, not role alternation, so it'd pass with an invalid output. Please add a role-alternation assertion (or merge adjacent user turns) and validate. Leaving open for you. 🙏

[diegosouzapw]

Merged — thank you, @KooshaPari! 🙏 To be clear, this is not subsumed by the earlier #6026 fix (#6115): that one strips orphan tool_result on the Antigravity→OpenAI MITM path, whereas this PR fixes the opposite direction (OpenAI→Claude) by reattaching a tool_result immediately after its matching tool_use when an intervening user turn separated them (Anthropic requires strict adjacency). I synced your branch onto the current release tip (the file-size baseline conflict was a pre-existing release-side rebaseline, resolved in favor of the release). Your toolResultAdjacency extraction + the regression test (17/17 green here) land as-is. Ships in the next release.