test(config): cover the four-layer config precedence chain

[acamarata]

src/config/mod.rs had zero tests. Adds 12 covering DaemonConfig::new, which layers CLI/env args > process environment > {data_dir}/config.toml > built-in defaults.

The mutants that matter here reorder an .or() chain, swap a default constant, or drop an .is_empty() filter — so every field is exercised at more than one layer. A test that only checks the default cannot tell a correct chain from one that ignores its overrides.

Environment safety

These mutate process-global state, and Rust runs tests in parallel threads. This follows the idiom already established in session::cursor::tests (a module lock held across the whole sequence, tolerating a poisoned mutex) and adds a Drop-based guard that clears all eight CLAWD_* variables on entry and restores them on drop, including on panic.

Without the clear, a "the default applies" test would fail on any machine that happens to export CLAWD_API_URL — an environment-dependent test is its own kind of hollow gate. Verified by running the full 706-test lib suite twice with zero failures.

Verification

Hand-applied six mutants; each fails a test:

Mutant	Result
DEFAULT_PORT 4300 → 8080	KILLED
default_bind_address → 0.0.0.0 (LAN exposure)	KILLED
port precedence flipped to toml.port.or(port)	KILLED
bind precedence putting env before the CLI arg	KILLED
drop the is_empty filter on CLAWD_REGISTRY_URL	KILLED
log default info → warn	KILLED

clippy --all-targets --all-features -- -D warnings and cargo fmt --check both exit 0.

One finding recorded, not changed

api_base_url and relay_url are the only two env reads without a .filter(|s| !s.is_empty()):

Variable set to ""	Result
CLAWD_API_URL	wins over the TOML value and blanks the URL
CLAWD_RELAY_URL	wins and blanks the URL
CLAWD_REGISTRY_URL, CLAWD_BIND, CLAWD_UPDATE_POLICY, CLAWD_LOG_FORMAT, CLAWD_LICENSE_TOKEN, CLAWD_API_TOKEN	falls through to the next layer

An unset shell variable expands to "", so CLAWD_API_URL=$SOME_UNSET_VAR silently points the daemon at an empty API URL. A test pins the current asymmetry so it cannot drift silently; I did not add the filter, since that changes startup behaviour for anyone relying on it. Worth a decision.