ci: split the wire round trips out of the fuzz marker

[ericeil]

The nightly Integration Tests job died in run template fuzz tests the first night the wire
round-trip suite existed — run 32218703320,
13 errors, one cause.

-m 'fuzz' tests selects two unrelated suites, and only one of them is Python. The 36 template
tests the step is named for, and the 13 round trips in tests/test_wire_roundtrip.py, whose
wire_echo fixture builds wire-echo with cargo at test time. The step runs pytest -n 2 and that
fixture is session-scoped per xdist worker, so both workers asked rustup to install the pinned 1.96
toolchain at the same instant. rustup's install path is not safe against concurrent invocations: one
call rolled back (failed to install component: 'rustc-x86_64-unknown-linux-gnu', detected conflict: 'bin/rust-gdb') and the other found what the rollback left behind (the 'cargo' binary ... is not applicable to the '1.96-x86_64-unknown-linux-gnu' toolchain). Not a flake that turned:
rust-toolchain.toml and the suite both landed in #97 on 08-18, after that morning's scheduled run,
so this was the first nightly to select them — 29 fuzz tests on 08-18, 49 on 08-19.

Installing the toolchain first would stop the crash and buy nothing. That is what pytest.yml
does, for exactly this hazard, and it is the wrong answer here: this job would pay an uncached cargo
build inside a 20-minute budget to re-run tests it cannot deepen and does not uniquely cover. They
pin max_examples=200 in their own @settings, so this job's HYPOTHESIS_PROFILE: extended never
reaches them — that profile is registered in test_fuzzed_templates.py alone. And pytest.yml
selects -m 'not expensive', which includes them, on every PR and every push to master, with the
toolchain installed and rust/target cached.

So the marker splits by requirement rather than by kind. wire says nothing about what those tests
do — they are fuzz tests — only about what selecting them commits a job to: having the pinned Rust
toolchain. A job with no Rust can now ask for fuzz and stay that way, which is what this one wants.
The sync-deps comment claiming the job "doesn't compile the Rust crates" stopped being true the
moment those tests carried fuzz; it now says that no toolchain gets installed, and points at the
step that depends on that.

Tests

The selector counts are the whole behavioural change:

-m 'fuzz'          → 36  test_fuzzed_templates.py   the nightly, Rust-free again
-m 'wire'          → 13  test_wire_roundtrip.py
-m 'not expensive' → 35  test_wire_roundtrip.py     pytest.yml, unchanged

pytest -m wire tests/test_wire_roundtrip.py: 13 passed. pytest -m "not expensive": 1036 passed,
11 skipped. pyright: 0 errors.

Nothing in PR CI exercises the nightly's selector — pytest.yml never runs -m fuzz — so the first
real proof is the next scheduled run.

🤖 Generated with Claude Code

[chandrakananandi]

Thanks for fixing this!

[chandrakananandi]

Seems like a better fix. Thanks!