feat(mdns): configurable skipInterfaces skip-list

[keskad]

Summary

• Adds a configurable skipInterfaces list to Config (JSON, default []). The built-in docker/veth/br-*/cni/flannel/virbr list is always applied; skipInterfaces adds extra case-insensitive name-prefix matches on top.
• wlan* is not skipped by default, so mDNS advertises on WiFi on a generic/laptop install. Operators who reserve the WiFi radio for another purpose add ["wlan"] (e.g. the BigFred hub, where wireless-programmer owns the radio, so mDNS must not leak bigfred.local / dcc-bus beacons onto a device config network).
• Threaded &[String] through should_skip_iface, preferred_ipv4_*, preferred_ipv6_addrs, preferred_iface_indexes, and MdnsPublisher::register, carrying the list via AnswerSet/DesiredAds so it flows from Config through run.rs and legacy_unicast.rs and picks up hot-reloads.

The matching bigfred-os change sets skipInterfaces: ["wlan"] in the hub's microdns.json.

Test plan

• cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test clean.
• skip_virtual_ifaces: wlan0/WLAN0 not skipped with &[], skipped with &["wlan"]; built-in virtual ifaces still skipped with &[].

[keskad]

Review — configurable skipInterfaces

The design is right: keep the container/virtual-bridge list hard-coded (it's never wrong to skip veth/docker/br-), and make the site-specific part configuration. Defaulting to [] so wlan* is advertised means a laptop/dev install keeps working, and only the hub opts out — which is exactly the shape this needed.

Verified locally:

• cargo clippy --all-targets clean, full test suite passes.
• #[serde(rename_all = "camelCase")] on Config makes "skipInterfaces" the correct key for the bigfred-os side (feat(os): skip wlan* in microdns on the hub bigfred-os#30), and #[serde(default)] keeps existing config files loading unchanged.
• Threading the list through DesiredAds and AnswerSet (both PartialEq) is the subtle part and it's done right: a hot-reloaded change to skipInterfaces now makes desired != last_desired and *w != next, so services are re-registered and multicast memberships refreshed rather than the change only taking effect on restart.
• The matcher tests cover the interesting cases, including that a blank entry matches nothing and that prefix matching doesn't accidentally catch wlp3s0 under ["wlan"].

1. should_skip_iface allocates on every call, per entry

skip.iter().any(|p| {
    let p = p.trim().to_ascii_lowercase();
    !p.is_empty() && n.starts_with(&p)
})

to_ascii_lowercase() builds a fresh String for every skip entry, for every interface, on every call — and the function is called from four preferred_* loops plus the interface poll every ifaceMs (5 s). It's small in absolute terms, but this repo is deliberately allocation-conscious and the inputs are constant between reloads.

Two ways out. Normalize once at load (trim, lowercase, drop empties in Config::validate), which also makes the stored config self-describing; or compare in place, since n is already lowercased:

skip.iter().any(|p| {
    let p = p.trim();
    !p.is_empty() && n.get(..p.len()).is_some_and(|head| head.eq_ignore_ascii_case(p))
})

get(..) rather than direct slicing avoids a panic if an entry's byte length lands mid-char.

2. run.rs:175 diagnostic no longer tells the whole truth

"no UP non-loopback IPv4 (skipping docker/veth/br-*)"

On the hub with skipInterfaces: ["wlan"] and Ethernet unplugged, this is precisely the message an operator sees, and it actively hides the reason — wlan0 has an address, it's just skipped by policy. Suggest appending the configured list, e.g. ... (skipping docker/veth/br-* + configured ["wlan"]). This is the one place where the new feature's failure mode is user-visible, so it's worth the extra formatting cost.

3. Nothing pins the JSON key

tests/mdns_test.rs covers the matcher thoroughly, but there's no deserialization test, so the wire name skipInterfaces is guarded only by the struct-level rename_all. bigfred-os#30 ships a file containing that exact key, and a future field rename would turn it into a silent no-op (no deny_unknown_fields). A couple of lines in the config test suite would close it:

let c: Config = serde_json::from_str(r#"{"services":[],"skipInterfaces":["wlan"]}"#).unwrap();
assert_eq!(c.skip_interfaces, vec!["wlan".to_string()]);
// and: a config without the key yields an empty list

4. refresh_memberships takes a second, independent read lock

let skip: Vec<String> = state.read().map(|g| g.skip_interfaces.clone()).unwrap_or_default();

This is a separate acquisition from the one that produced want_v4, so in principle the addresses and the skip list can come from different generations of the AnswerSet. Harmless — the next poll converges, and a torn read here only costs one refresh cycle — but folding it into the existing guard would remove the question entirely.

5. has_usable_iface / primary_ip parameter churn

Both gained a skip: &[String] parameter but appear to have no in-tree callers. Fine if they're deliberately part of the public surface; otherwise they're signature churn on dead code and could keep their old shape (or go away).

6. Prefix matching is a foot-gun worth documenting

skipInterfaces: ["e"] silently disables eth0 and enp1s0, and the only symptom is "no UP non-loopback IPv4". The README addition is good on the why, but a one-line caution that entries are prefixes (not names or globs) — or a validate() warning for very short entries — would prevent someone reaching for ["w"] and losing mDNS entirely.

[keskad]

Review points applied

Pushed as 27a434a:

• should_skip_iface no longer allocates. The name is already lowercased, so the head is compared with eq_ignore_ascii_case via get(..len) — no String per skip entry per interface per poll, and no panic risk on a byte length landing mid-char.
• refresh_memberships takes one read lock, so the addresses and the skip list are guaranteed to come from the same generation of the AnswerSet.
• The "no UP non-loopback IPv4" diagnostic names the configured skips. This is the message a hub operator sees when Ethernet is unplugged and wlan0 is the only addressed interface, and it previously hid the reason completely.
• Two serde tests pin the JSON contract: "skipInterfaces": ["wlan"] deserializes into the field, and a config without the key still yields an empty list so existing files keep advertising on wlan*. bigfred-os#30 ships that exact key and unknown fields are ignored, so a rename would otherwise silently turn the hub's opt-out into a no-op.
• README spells out that entries are prefixes, including the ["e"] foot-gun that would take eth0/enp1s0 with it.

cargo clippy --all-targets clean, 40 tests passing (was 38).

Left alone deliberately: has_usable_iface / primary_ip gained the parameter but have no in-tree callers — they were already unused public functions before this PR, so changing their shape here would be unrelated churn.