Carry the runtime Context into thread pool workers

[tamohannes]

make test is red on main: 5 failures, all the same error.

LookupError: <ContextVar name='nu_rt_ctx'>
  src/nu/lang/runtime/runtime.py:75  in Runtime.ctx

Cause

Runtime.ctx resolves through the _RT_CTX ContextVar, which the comment above it describes as per asyncio task with copy on write inheritance. That holds for tasks. It does not hold for threads: a ThreadPoolExecutor worker starts with an empty contextvars context, so _RT_CTX is simply unset there.

Every hand-off to the Budget's pool goes over that boundary bare:

futures = [pool.submit(self.eval, n) for n in nids]          # eval_parallel
return await loop.run_in_executor(pool, self.eval, n)        # _drive_async

So a parallel branch dies as soon as it touches the Context, which is what ref._write does on the way into ctx.attrs.

Regression from 322f9f3 ("Make Runtime.ctx per-asyncio-task via ContextVar"), 2026-07-21.

Fix

Take a fresh contextvars.copy_context() on the calling side and submit through its run. Six dispatch sites: eval_parallel, _drive_async, merge, amerge, in_thread, a_in_thread.

A copy rather than the caller's own context is deliberate, and it gives exactly the semantics the module comment already claims: the Context object is shared so writes to ctx.attrs still land where the caller can see them, while rebinding _RT_CTX inside a branch stays local to that branch. It also avoids the "cannot enter context: already entered" case, since two workers never share one Context object.

Verification

Before: 5 failed, 2053 passed. After: 2058 passed, 27 skipped.

Added one regression test per dispatch site under a new "context across the thread boundary" section in tests/nu/lang/runtime/test_runtime.py. merge, amerge, in_thread and a_in_thread had no coverage for this at all. All six fail on main and pass here.

Found while checking #2 on a clean clone, so this is what make test does out of the box, not a code read.

🤖 Generated with Claude Code