test(config): serialize test_from_figment_cli_args_override

[wagenet]

config::tests::test_from_figment_cli_args_override was the only test in the
config module without #[serial]. It fails intermittently in a full parallel
cargo test run with:

called `Result::unwrap()` on an `Err` value: Invalid argument (os error 22)

Cause

Every neighbouring test_from_figment_* test moves the process working
directory into a TempDir with std::env::set_current_dir, then lets the
TempDir drop. #[serial] serializes those against each other but not against
this one, so it can run inside that window.

Config::from_figment looks for .strom.toml in the process working directory.
When it runs while test_config_file_with_data_dir is chdir'd into its
TempDir, it reads that test's config file, picks up its
storage.data_dir — which points inside the same TempDir — and
DataPaths::resolve calls create_dir_all on a directory the other test is
deleting. On macOS that mkdir mostly fails with EINVAL, which is the
reported os error 22; occasionally ENOENT.

Note this is not the working directory read failing. from_figment calls
std::env::current_dir().ok(), so an unlinked working directory is swallowed
(and would give ENOENT, not EINVAL). The failure comes from the config file
the borrowed working directory exposes.

Confirmation

Forced the race with a throwaway probe (not committed): a background thread
looping chdir-into-TempDir / write .strom.toml / delete, while the main thread
ran this test's from_figment call 4000 times. Three runs:

run	failures / 4000	EINVAL	ENOENT
1	214	204	10
2	219	210	9
3	210	201	9

That reproduces the exact reported message.

Fix

Add #[serial], matching every other test in the module.

The two remaining unserialized tests in the module are not exposed:
test_ice_servers_normalization is a pure string test, and
test_legacy_config_new calls Config::new with data_dir: None and no
config-file or environment read, so its base directory is always the platform
data directory.

Testing

• cargo test --features efp --lib, 12 runs before the change and 12 after — all
  544 tests green in every run. A green run is weak evidence: the flake is
  reported at roughly 1 in 8 full runs, so 12 clean runs before the change is a
  ~20% likely outcome on its own, and I did not observe the flake in the baseline.
  The argument for this fix is the serialization reasoning plus the forced-race
  reproduction above, not the run count.
• No deterministic regression guard is included. A committed version of the probe
  would prove from_figment is working-directory-sensitive, but it would not fail
  if the #[serial] were reverted, so per the repo's test rules it would be a
  demonstration rather than a guard. Forcing the real race would require a thread
  mutating the process working directory outside the serial lock, which is the
  disease, not a test.

🤖 Generated with Claude Code