fix(core): don't fail sibling tools on tool-output encode errors

[xiaomayi-ant]

Issue for this PR

Closes #47024

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

I was reading through session/runner/llm.ts to understand how tool execution works in the V2 runner, and got stuck on the raceFirst(FiberSet.join, FiberSet.awaitEmpty) logic in awaitToolFibers. What confused me: if the model dispatches several tool calls at once and just one of them fails, the runner ends up marking all the other still-running ones as failed too — even ones that would've finished fine on their own. That felt wrong, so I dug into what could actually make join fail, and it turns out the only thing typed to reach it is ToolOutputStore.Error — a StorageError tagged with either operation: "encode" or "write".

That's basically the whole fix: an encode failure only means that one tool's output couldn't be serialized, it has nothing to do with anyone else's storage — so there's no good reason to fail the whole batch because of it. I added a short extra wait (FiberSet.awaitEmpty) before sweeping unsettled tools in that specific case, so the other calls get a real chance to finish and report their own actual result instead of getting overwritten with a false "failed". A write failure is more likely a shared disk/storage problem, so that path still fails fast like before — didn't touch it.

How did you verify your code works?

• bun run typecheck in packages/core passes; bun turbo typecheck across all 30 packages also passes (ran automatically via the pre-push hook).
• bun test in packages/core: 1096 tests pass, no regressions.
• Didn't add a new test for this specific race (concurrent tool calls + one encode failure) — flagging that so a maintainer can weigh in on whether one's expected.

Screenshots / recordings

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR