fix(rate): initialise a stat that is new since the previous sample - #3721

compute_rate_on_list() looks up each stat's previous sample by key and skips the stat when there is none:

for stat in stats:
    old = previous_by_key.get(stat[key])
    if old is not None:
        compute_rate(self, stat, old)

The dict path treats a missing previous as a first sample and initialises the stat. The list path skipped it. Interfaces, disks and VMs appear while Glances is running — a VPN comes up, a container starts, a disk is plugged in — so this is not only a start-up case. network, diskio and vms all carry rate fields on a list.

A skipped stat keeps the raw psutil counter in the delta field and gets no _gauge, no _rate_per_sec and no time_since_update:

round 2  eth0  {'bytes_recv': 0, 'time_since_update': 3.0e-05, 'bytes_recv_gauge': 1500, 'bytes_recv_rate_per_sec': 0}
round 2  wg0   {'bytes_recv': 8000000000}

Two consequences

The refresh publishes the lifetime total as one refresh of traffic. bytes_recv is documented as the delta, and for one sample it holds the since-boot counter instead. On a quiet link that first sample reads in gigabytes.

NetworkPlugin.update_views() raises. It assigns views/<iface>/bytes_recv_rate_per_sec directly, and update_views() in the base class builds views only from listkeys(i) — the fields a stat actually carries. With no rate field there is no view to decorate:

glances/plugins/network/__init__.py:237: KeyError: 'bytes_recv_rate_per_sec'

Driving the real plugin from a fake psutil.net_io_counters, one interface appearing is enough to reach it. diskio happens to survive because it guards with if f'{field}_rate_per_sec' in disk_views; network does not.

The change

One line: pass an empty previous rather than skipping.

compute_rate(self, stat, previous_by_key.get(stat[key], {}))

compute_rate already handles an empty previous — it is the branch a dict stat takes on its own first sample. The gauge is recorded, the rate reads 0, and the next sample measures a real delta. That also gets the stat to its second sample one refresh sooner, since the gauge it needs now exists the moment it appears.

I did not add a guard in network/update_views() instead, because the missing views are a symptom: the stat was left in a shape no other stat is ever in, and vms and diskio were publishing wrong numbers from the same cause without raising.

Verification

tests/test_rate_on_list.py, 5 tests: four on the shared rate code through a minimal list plugin, one end-to-end through NetworkPlugin for the KeyError.

Checked by breaking it — reverting the one changed line back to the skip:

Full suite on Windows, before and after: the set of failing test names is byte-identical (63 pre-existing environment failures — the xmlrpc and browser-restful servers). ruff check and ruff format --check clean.