Fix container network accounting: sum all interfaces, not only eth0 #3669

Context

DockerStatsFetcher._get_network_stats() in
glances/plugins/containers/engines/docker.py reads network counters from a
single hardcoded interface:

eth0_stats = self._streamer.stats.get('networks', {}).get('eth0')

The Docker stats endpoint returns one entry per container interface. A container
attached to several Docker networks (docker network connect, or several
networks declared in a Compose file) gets eth0, eth1, … and everything past
eth0 is silently dropped. Reported RX/TX rates for those containers are too
low.

Verified on both develop and develop-v5: the file is byte-identical on
the two branches, so the same patch applies to both.

Scope

In scope: _get_network_stats() in
glances/plugins/containers/engines/docker.py, plus unit tests.

Out of scope — do not touch:

• glances/plugins/containers/engines/podman.py. Podman exposes only a
  pre-aggregated NetIO string; there is no per-interface data to sum and no
  bug to fix there.
• The containers plugin model, the TUI, or the WebUI. The output contract of
  _get_network_stats() is unchanged: same keys, same types, same units.
• Any adjacent code, comment or formatting in docker.py (CLAUDE.md §3).

Non-regression requirements

These must all hold after the change:

1. Single-interface containers (the majority) produce identical values to
   the current implementation.
2. Containers started with --network host have no networks key in the stats
   payload. The method must keep returning None, exactly as today.
3. The returned dict keeps the same shape: cumulative_rx, cumulative_tx, and
   — only when previous stats exist — time_since_update, rx, tx.
4. No new dependency, no new configuration key.

Plan

1. Patch `_get_network_stats()` to aggregate across interfaces
   → verify: `make lint && make format` clean; the diff touches only the
     interface-selection and summation lines of that one method
 
2. Add `tests/test_plugin_containers.py` covering single-interface,
   multi-interface, host-network, malformed-payload and rate-computation cases
   → verify: `uv run pytest tests/test_plugin_containers.py` passes
 
3. Run the existing suite to confirm no collateral damage
   → verify: `make test` shows no new failures versus the pre-change baseline
 
4. Add a NEWS.rst entry
   → verify: entry follows the existing `.rst` bullet format under the
     unreleased section

Step 1 — the patch

Replace the interface selection and the stats dict construction. Keep the rest
of the method (rate computation from _old_computed_stats) untouched.

        interfaces_stats = self._streamer.stats.get('networks') or {}
 
        # Checks for net_stats & mandatory fields
        # Note: a container can be attached to several networks (eth0, eth1...)
        # so all the interfaces should be aggregated (issue #XXXX)
        valid_interfaces = [
            interface_stats
            for interface_name, interface_stats in interfaces_stats.items()
            if interface_name != 'lo' and all(field in interface_stats for field in ['rx_bytes', 'tx_bytes'])
        ]
        if not valid_interfaces:
            self._log_debug("Missing Network usage fields")
            return None
 
        # Read the rx/tx stats (in bytes), summed over all the container interfaces
        stats = {
            'cumulative_rx': sum(interface_stats['rx_bytes'] for interface_stats in valid_interfaces),
            'cumulative_tx': sum(interface_stats['tx_bytes'] for interface_stats in valid_interfaces),
        }

Replace #XXXX with the real issue number before committing.

Two deliberate choices, do not silently change them:

• lo is excluded. The Docker daemon does not normally report loopback in
  networks, but Docker-API-compatible runtimes (nerdctl/containerd) have been
  observed to. Loopback traffic would massively inflate the totals, so the guard
  is cheap insurance rather than defensive noise.
• Interfaces with missing counters are skipped, not fatal. Previously a
  malformed eth0 aborted the whole method. Now one malformed interface among
  several no longer discards the valid ones. None is still returned when
  nothing usable remains, which preserves requirement 2 above.

Decision required before implementing — do not decide alone

Summing across interfaces introduces a case that did not exist before: if a
container is attached to or detached from a network between two samples, the
interface set changes and cumulative_rx/cumulative_tx jump or drop. A drop
produces a negative rx/tx delta for one cycle.

Note that negative deltas are already possible today on container restart
(counters reset), so this is an aggravation of an existing behaviour rather than
a new class of bug.

Options:

• (a) Leave as is. Strictly surgical; one bogus sample in a rare scenario.
• (b) Clamp to zero: max(0, stats['cumulative_rx'] - old[...]). Hides the
  glitch, but also hides genuine counter resets and changes existing behaviour
  for the restart case.
  Default: (a). Do not implement (b) without explicit approval.

Step 2 — tests

Create tests/test_plugin_containers.py.

Style note: tests/test_plugin_sensors.py is written in pytest style
(fixtures, plain classes, bare assert) while older files such as
tests/test_core.py use unittest.TestCase. Match the sibling
test_plugin_*.py convention, i.e. pytest style, per CLAUDE.md §3.

DockerStatsFetcher.__init__ opens a live stats stream, so do not call it.
Build the instance without the constructor and inject the two attributes the
method reads:

from types import SimpleNamespace
 
from glances.plugins.containers.engines.docker import DockerStatsFetcher
 
 
def build_fetcher(networks, old_stats=None):
    fetcher = DockerStatsFetcher.__new__(DockerStatsFetcher)
    fetcher._container = SimpleNamespace(id='deadbeef')
    fetcher._streamer = SimpleNamespace(
        stats={'networks': networks} if networks is not None else {},
        last_update_time=100.0,
    )
    fetcher._old_computed_stats = old_stats or {}
    fetcher._last_stats_computed_time = 97.0
    return fetcher

Cases to cover:

Step 3 — regression run

Capture a baseline before touching anything:

make test 2>&1 | tail -30   # baseline
# ... apply the change ...
make test 2>&1 | tail -30   # compare

Only the new test file may add results. Any other delta is a regression, stop
and report it.

Step 4 — changelog

Add to NEWS.rst, under the unreleased section, bug-fix list:

* Container network stats now aggregate all the container interfaces
  instead of eth0 only (issue #XXXX)

Branches

Apply to develop first. develop-v5 currently holds an identical copy of the
file, but re-verify the divergence before assuming a clean cherry-pick —
confirm the method body still matches, then port.

Report back

• Diff of docker.py (should be roughly 10 lines).
• Test file, and its pass output.
• Baseline vs post-change make test comparison.
• Your recommendation on the decision point above, with reasoning.