APuP/batman-adv: address #1200 review (fail-fast primitives, limed guard, cleanup)

[Fede654]

Follow-up on libremesh#1200 picking up the pending review changes (per @javierbrk's "if anyone wants to pick it up in the meantime, that's ok with me"). This PR targets the fix/ap-up_batman_mac branch so it folds back into libremesh#1200.

Each commit maps to a specific review request and keeps the design criteria already agreed in the thread (fail-fast primitives, defensive handling only at the call site, reuse of existing idioms, no leaked globals / dead code). All unit tests pass (./run_tests): 305 successes / 0 failures / 0 errors, unchanged count vs. the current head.

Changes

1. utils.split fails on an empty separator (commit 1)
Replaces the silent whitespace default with assert(sep ~= nil and sep ~= '', ...), and drops the nil/empty-string guard that was only there to absorb get_mac returning nil (now fixed at its source). Audited all 18 call sites: every one passes an explicit non-empty separator, and gmatch already yields {} for an empty input string, so valid callers are unaffected.

Addresses: "empty separator should be treated as error, the function should fail in that case because calling it without separator is symptomatic of something wrong in the caller."

2. network.get_mac fails fast; device_exists via stat (commit 2)
get_mac now asserts when the interface has no MAC (does not exist) instead of returning nil, surfacing the caller bug with a clear message. The NAK'd test is updated to assert.has_error. Separately, network.device_exists is reimplemented with fs.lstat("/sys/class/net/"..dev) — the same lookup already used by network.assert_interface_exists and get_own_macs — instead of io.popen("ip link show") + stdout parsing, since it runs on the hot path of every peer notification.

Addresses: "better to fail ASAP if the MAC address for a non-existent device is requested, instead of propagating a nil value upstream" and the NAK on the nil-returning test.

3. limed: skip a missing APuP peer before configuring it (commit 3)
Moves the device_exists guard ahead of network.createStatic, since createStatic is what shells out to ip link set up / ip address add and produced the No such device errors. The verbose pasted-log comment is replaced by a concise explanation: this is an expected transient during a wifi down/wifi up cycle (the APuP peer interfaces are not torn down in sync and reappear on re-association), not an error. Also fixes the skip log line, which passed the ifname as a second arg to a string.format call with no specifier, silently dropping it.

Addresses the question on whether the missing-interface condition is normal or an error: it is a normal transient, now handled and documented as such.

4. lime-proto-batadv: scope the per-peer MAC to a local + cleanup (commit 4)
macaddr was assigned without local, leaking a global; scoped to the function. The derivation is unchanged and now documented inline: 02 : id[2] : id[3] : radio[4] : radio[5] : radio[6] (locally-administered unicast; two bytes from the md5-derived interface id, low three bytes from the radio MAC — unique per peer, keeps node identity). utils.log uses a format string instead of concatenation, and the commented-out batctl if add dead code is replaced with a one-line rationale.

Design note

The review asked for fail-fast in the primitives; the defensive nil-tolerance existed because of the wifi down/up race. These reconcile by keeping the primitives strict and putting the tolerance only where the transient is legitimate — the device_exists guard in limed — which is exactly what makes the downstream fail-fast safe. No robustness is lost.