feat: sparkDash-derived host + serving metrics for every gateway

Phase 1/2 of the sparkDash integration, gateway-agnostic end to end
(all data flows through the existing agent payloads; no daemon, no
web server):

Agent 1.2.0 (agent_core collectors, pure parsers, graceful None):
- uptime (proc/uptime or kern.boottime), storage (statvfs), network
  rx/tx rates (/proc/net/dev deltas, loopback/veth excluded)
- tailnet self-health (Self.Online, backend state, health warnings -
  the healthy-on-LAN-but-off-tailnet blind spot), TTL-cached
- GPU process names via nvidia-smi --query-compute-apps
- live serving rates for running models: llama.cpp /slots per-slot
  counter diffs, vLLM Prometheus generation/prompt token counters
  (+ KV cache %, running/waiting), sglang /server_info counters with
  sticky-throughput expiry. TTL-coalesced, loopback-only unless
  ALLOW_REMOTE_HEALTHCHECK, idle reads 0, all failures degrade.

Swift (additive decode, old agents tolerated):
- HostMetricsPayload: uptime_seconds/storage/network/tailscale,
  HostGPUProcess.name; ModelProfileStatus.serving (ServingMetrics)
- Remote Hosts card: uptime in header, STORAGE/NET lines, tailnet
  status dot with warning reasons, GPU process name under models,
  tok/s on running rows; dashboard row subtitles prefer live tok/s
  over stored benchmark rows

Tests: 21 new collector/parser tests (failure-first), 5 decoder/
presentation tests, conformance allowlist updated for .
Verified collectors live on Spark (uptime 1.0h, storage 71.7%,
net 66.7 kbps, tailnet Running/100.122.96.76). 187 swift, 57 python.

Author: AdityaVG13

CHANGELOG.md

@@ -3,6 +3,11 @@
 All notable changes to this project are documented in this file.
 
 ## [Unreleased]
+### Added
+
+- sparkDash-derived host visibility for every remote gateway (agent 1.2.0): uptime on the gateway header, storage and network rate readouts, tailnet self-health (online state, backend state, warning reasons - catches the "healthy on LAN but off-tailnet" blind spot), and GPU process names next to running models.
+- Live LLM serving rates on model rows and the Remote Hosts panel: decode/prefill tok/s from llama.cpp `/slots`, vLLM Prometheus `/metrics`, and sglang `/server_info` counters, diffed over time so idle reads 0; plus KV-cache usage and request queue depth for vLLM. Loopback-only unless `ALLOW_REMOTE_HEALTHCHECK`, same SSRF posture as health probes; every probe failure degrades to a hidden label, never an error.
+
 ### Fixed
 
 - Gateway agent updates can no longer hang forever on "Pushing agent…": each deploy step is killed at a 120s deadline with a pointed message, and Tailscale SSH re-auth prompts ("To authenticate, visit https://login.tailscale.com/…") are classified into remediation instead of stalling BatchMode ssh silently.