L2 dispatcher passes xdist options to a child with xdist disabled

[doraemonmj]

Platform

a2a3sim (the command-construction bug is platform-independent)

Runtime

Multiple L2 runtimes in one pytest invocation, for example host_build_graph and tensormap_and_ringbuffer

Description

When the top-level pytest invocation disables xdist with -p no:xdist, the L2 phase dispatcher inherits that option into each child command. If the effective --max-parallel value is greater than 1 and the xdist package is importable, the dispatcher also appends -n <N> --dist loadfile.

The resulting child command both disables xdist and uses xdist-only CLI options. Pytest then exits during argument parsing with rc=4.

This is easy to miss because a collection containing only L2 items from one runtime takes the direct-run fast path and does not invoke the L2 dispatcher. A collection spanning multiple runtimes does invoke it and fails deterministically.

Steps to reproduce

From the repository root with the test dependencies installed:

python -m pytest \
  tests/st/a2a3/tensormap_and_ringbuffer/dfx/args_dump/test_args_dump.py \
  tests/st/a2a3/host_build_graph/vector_example/test_vector_example.py \
  --platform a2a3sim \
  --device 0-1 \
  -p no:xdist \
  --max-parallel 2 \
  --manual include \
  --pto-session-timeout 600

No device allocation is required to observe the failure because both child processes fail while parsing arguments.

Expected behavior

The dispatcher must not construct a child command that disables xdist while also passing xdist-only options.

The intended semantics of -p no:xdist should be made explicit. Either:

1. Treat it as authoritative for all phases and run L2 children serially without -n/--dist; or
2. Treat it as applying only to the outer orchestrator and remove only the no:xdist plugin-disable option from L2 child arguments before adding -n/--dist.

Actual behavior

Both runtime children fail before collection:

::group::L2 host_build_graph [-n 2]
ERROR: ... unrecognized arguments: -n 2 --dist loadfile
--- L2 host_build_graph: FAIL rc=4 ---

::group::L2 tensormap_and_ringbuffer [-n 2]
ERROR: ... unrecognized arguments: -n 2 --dist loadfile
--- L2 tensormap_and_ringbuffer: FAIL rc=4 ---

Root cause

_base_pytest_argv() in conftest.py inherits the original invocation arguments. The L2 path only strips --exclude-level, so -p no:xdist remains in base_args.

_dispatch_test_phases() independently checks whether import xdist succeeds when max_parallel > 1. It does not check whether the plugin was explicitly disabled, and then appends -n <N> --dist loadfile.

Importability is therefore being used as a proxy for plugin availability in the child, but those are not equivalent when -p no:xdist is inherited.

Suggested regression coverage

Add a dispatcher command-construction test with:

• multiple L2 runtimes so the dispatcher path is taken;
• -p no:xdist in the inherited invocation arguments;
• --max-parallel 2; and
• pytest-xdist installed.

Assert that a child command never contains both no:xdist and -n/--dist, and add an end-to-end subprocess assertion that the invocation does not exit with pytest usage error rc=4.

Commit

58ca976f97c7ac9461f8c0ae8f2db3b7fe4f6f78

Host platform

Linux aarch64, Python 3.10.19, pytest 8.4.2, pytest-xdist 3.8.0