🐛 Benchmark suite: `test_querying.py` mocks `NeedsSphinxConfig` without `variant_data_proxy` (12 failures CI never collects)

[chrisjsewell]

What happens

uv run poe benchmark-needs (or pytest tests/benchmarks) fails 12 of the 18 tests in
tests/benchmarks/test_querying.py:

FAILED tests/benchmarks/test_querying.py::test_filter_single_need_simple[100]
…
FAILED tests/benchmarks/test_querying.py::test_filter_needs_and_parts_complex_eval[1000]
E   AttributeError: Mock object has no attribute 'variant_data_proxy'.
    Did you mean: 'variant_data_file'?
12 failed, 6 passed

Reproduce quickly (no timing) with:

uv run --frozen --group benchmark pytest tests/benchmarks/test_querying.py --benchmark-disable -q

Why

test_querying.py builds its config as Mock(spec=NeedsSphinxConfig) (line 113). A spec'd
mock only has the attributes the class declares, and variant_data_proxy is not a field of
NeedsSphinxConfig: sphinx_needs/needs.py assigns it dynamically in
_derive_variant_data_proxy() (config.variant_data_proxy = …), and the filter code reads it
(filter_common.py, directives/needif.py). The dynamic assignment arrived with
needs_variant_data in #1715 (2026-05-26); the benchmark test has not changed since #1677
(2026-03-31), so the suite has been red since then.

CI does not see it: the benchmark workflow runs the directory twice with -k _time and
-k _memory, which together collect 3 of the 25 tests, all outside the failing set. It was
found by running every poe task on the workspace-layout branch (#1839), where the failure
reproduces identically on the moved tree and on master.

Possible fixes

• Declare variant_data_proxy on NeedsSphinxConfig (it is a derived value, so a field with
  a None default that _derive_variant_data_proxy fills), which also makes it visible to the
  type checker instead of being an untyped attribute; or
• keep it dynamic and have the benchmark fixture set it on the mock
  (config.variant_data_proxy = None).

Separately: should the benchmark job collect the whole directory, so the non-timed tests in
it are at least executed? Today a benchmark that stops working is invisible until someone
runs the task locally.