traffic dashboard: serve interface-traffic panels from gNMI end-to-end, stop using the InfluxDB fact table

[thijsvanemmerik]

Problem

The traffic dashboard panels (data.doublezero.xyz) still depend on InfluxDB. On lower / zoomed-in timeframes (sub-5-minute buckets) they query the InfluxDB-derived fact_dz_device_interface_counters, which trails real time by ~1 hour, so those views show a trailing ~1h empty gap on every device. Now that every device emits gNMI, the panels should be served from gNMI end to end and stop using InfluxDB.

Evidence

Freshness by table (mainnet-beta, 2026-07-18):

Table	Fed by	Lag
telemetry_<env>.interface_state (raw gNMI)	gNMI tunnel	~0 min (real-time)
device_interface_rollup_5m	gNMI (interface_state)	~3 min
fact_dz_device_interface_counters	InfluxDB-derived	~64 min

• Panel routing in api/handlers/traffic_dashboard.go (isRawBucket / trafficTimeFilter): sub-5m buckets use the raw InfluxDB fact table; >= 5m buckets use the gNMI rollup. So zoomed views fall onto InfluxDB and inherit its ~1h lag.
• gNMI coverage is complete: 97/97 current mainnet devices present in interface_state (24h), and zero current devices are InfluxDB-only (checked dz_devices_current against both sources). The InfluxDB fallback is no longer exercised.

What to do

1. Sub-5m / raw traffic path: read the gNMI interface_state table (real-time, ~30s samples, finer than 5m) with on-the-fly per-sample deltas, mirroring ComputeDeviceInterfaceRollupFromGNMI in indexer/pkg/rollup/activities.go, instead of fact_dz_device_interface_counters.
2. Rollup path: the gNMI-over-InfluxDB coalesce (coalesceDeviceInterfaceBuckets) already prefers gNMI; drop the InfluxDB fallback (or leave it inert) once coverage is confirmed.
3. Remove InfluxDB from the traffic panel query path entirely, so every zoom level is real-time gNMI.
4. Guard: before removing the fallback, add/confirm a check that every activated device has gNMI coverage (per-device coverage check, infra #1464), so a future gNMI-less device is caught rather than silently blank.

Scope note

fact_dz_device_interface_counters (InfluxDB-derived) has other consumers beyond these panels (traffic / topology / incidents / users handlers). This issue scopes to the traffic panels being served from gNMI. Migrating the remaining consumers and fully decommissioning the InfluxDB ingestion + backfill is a larger follow-up worth tracking separately.

Outcome

All interface-traffic panels served from gNMI (interface_state for sub-5m, device_interface_rollup_5m for >= 5m), real-time at every zoom level, with no InfluxDB in the panel path.