[timmy21]

Problem

gremlins left many mutants alive in cpdaemon/pkg/cpm and cpgolib/cpworker (#243 item 6): the tests run these branches but never check their result. Re-measured on 0.9.x (c4e1a8e9), test efficacy was 63.4% for pkg/cpm (79 surviving mutants) and 76.0% for cpgolib/cpworker (6 surviving). Among the gaps:

• which VXLAN tag format (vni1/vni2) a CPM strategy produces, and the 0xFFFFFF fallback without a service tag;
• how the NIC buffer and pipeline buffer are sized from the memory limit;
• what happens when CPM returns a new daemon id or the strategy sync keeps failing;
• whether a rejected control-socket handshake is noticed by the client.

Change

Tests only, no product code changes.

• worker_task_builder: the VXLAN tag format follows apiVersion (docs/VXLAN-WIRE-FORMAT.md §5, §8, §9): vni1 carries the service tag, or 0xFFFFFF without one; vni2 carries domain, point and extension flag, falling back to domain 1 and point 1. A rate limit of 0 is left unset. The vxlan port, zmq service tag and file dump interval reach the output. packetSplitBytes 0 is accepted. --nofilter without bind_device is rejected.
• worker_mgr: auto_nic_buffer splits the memory limit over the tasks. CPM memLimit overrides the default only when > 0. Too little memory is an error. The pipeline section gets the memory left after the libpcap buffers, but at least MinBufferSizeMb. cpu_affinity is passed through. The control socket directory is created before the worker starts.
• syncer: register sends the start time as seconds plus microseconds. A new daemon id stops the worker and resets the sync state. A failing strategy sync is retried SyncStrategyMaxRetries times. The metrics sampling microseconds stay below 1e6. The uuid file is reused, and an empty one is refilled. IncludingNICs limits the registered interfaces.
• client: SetDefaults keeps values that are set. Register sends labels and networkInterfaces as [], and fails on a non-2xx status or a body code >= 400.
• models, utils, synclog: HasInstances; splitArgs drops empty arguments; the log level threshold; log entry microseconds.
• cpgolib/cpworker: Dial fails when the handshake reply is not OK or not JSON. A timeout of 0 falls back to DefaultTimeout.

Not changed, as discussed in #243:

• Error branches that only log, CPU/memory figures read from the live process, and log-only formatting: low value, or nondeterministic.
• Omitting an empty arguments field in cpgolib requests: cpworker never reads arguments (unix-manager.c), so a test would pin serialization without a behaviour behind it.

Tests

gremlins v0.5.0 test efficacy, before → after:

• cpdaemon/pkg/cpm: 63.4% → 83.3% (measured before the last vni2 fallback case). The rest are the categories above plus equivalent mutants.
• cpgolib/cpworker: 76.0% → 85.7%. The rest are 3 equivalent mutants plus the arguments case.

gremlins copies the module to a temporary directory, which breaks cpdaemon's relative replace => ../cpgolib: every mutant then fails to compile and is counted as killed. The numbers above come from a copy with an absolute replace.

Suites run:

• go vet ./... and go test ./... for cpdaemon and cpgolib on macOS: pass.
• pkg/cpm and cpgolib/cpworker on Ubuntu 24.04: pass.

Refs #243