Push test coverage from ~94% to ~98%

[defcron]

Summary

This branch was originally opened as PR #3 stacked on the (then-open) PR #2. #2 has since merged into main, but #3 was merged into the now-gone binda-scaffold-and-features branch rather than main directly, so this coverage commit never made it into main. This PR re-targets that same work at main.

Closes nearly all remaining coverage gaps identified via cargo llvm-cov --show-missing-lines after #2:

• api.rs: converted match { ... => {} other => panic!(...) } test assertions to assert!(matches!(...)) — the panic! arm was itself an uncoverable dead branch on a passing run, which was inflating the apparent gap. Also added the one real gap this revealed: Register with a bad signature was never exercised.
• collision.rs / dns.rs / store.rs / gossip.rs: covered the remaining branch of LowerWins collision resolution, NS records, a sixth registration attempt while still live (not just after going stale), and made the ConformanceChallenge mismatch test deterministic instead of depending on which way a single random challenge happened to fall.
• fcrdns.rs / ntp.rs: added tests for is_forward_confirmed's two failure-propagation branches, AAAA forward lookups, a wrong-record-type filter, multi-resolver fallback, a compression pointer loop, malformed/short responses, and a short NTP reply — via the same local fake-server pattern already established in Behavioral gossip conformance testing, CI/coverage tooling, and ~94% test coverage #2.
• main.rs: extracted spawn_listener_tasks out of main() so the listener-spawning logic (which address goes to which listener, a --dns flag producing a fourth task, and each task's bind-failure logging) is unit-testable on its own. Excluded main() itself from coverage accounting with the nightly #[coverage(off)] attribute, since it's pure wiring around already-covered pieces that, by design, never returns in a real run.
• node.rs: added a smoke test for the previously-never-called spawn_rate_limiter_maintenance.

Workspace test count: 147 → 167. Overall coverage: ~94% → ~98%.

What's left, and why

The remaining ~2% is essentially irreducible without contrived tests:

• A handful of node.rs lines are the "kick off an infinite listener loop" statements inside tests that intentionally never let that loop finish, so the async state machine's completion path is never counted — even though the loop's actual behavior is thoroughly exercised via real socket round-trips elsewhere in the same test.
• A couple of fcrdns/ntp lines are a test-helper's own defensive error arm that only fires if a background thread's socket read fails after the test already got what it needed.
• A few are error variants (like a JSON encode failure) with no reachable trigger through this project's own types without constructing a deliberately-broken value.

Notes for reviewers

• #[coverage(off)] requires #![feature(coverage_attribute)], which is fine since the project already pins Rust nightly.
• All new tests verified stable across repeated runs (both --test-threads=1 and default parallelism).
• This is the same diff previously reviewed/merged as Push test coverage from ~94% to ~98% #3 (into the interim branch) — no new changes since then.

🤖 Generated with Claude Code