Telemetry capture gaps/wishlist for per-vCPU overhead calculations

While characterizing per-vCPU overhead on a rack, I tried to account for every lost cycle from telemetry alone. vcpu_usage, sled_cpu:cpu_nsec, and zone:cpu_nsec got ~90% there, retroactively and with no host access. The gaps below forced on-sled kstat/prstat or left numbers unmeasurable, ordered by impact:

1. Global-zone CPU attribution. zone:cpu_nsec reports the global zone as a single bucket, so platform work inside it (viona, xde, ZFS, kernel net) is indistinguishable. Coarse kthread-group attribution would close the biggest gap.
2. Effective CPU frequency (APERF/MPERF). Per-vCPU compute swings weren't visible in any counter. Clocks are cheap to sample and would expose the frequency share. (Memory-bandwidth counters are similar but lower priority.)
3. Split emulation into kern/user. Kstat has time_emu_kern / time_emu_user; vcpu_usage collapses them. In-kernel vs propolis-userspace matters for optimization targeting. This is already collected, just merged.
4. OPTE/xde port metrics. Nothing between instance_network_interface and sled_data_link: softring drops, flow/route-cache occupancy, encap throughput would be helpful. Networking investigations needed on-sled kstat/opteadm. Viona ring stats likewise.
5. Crucible latency below the guest view. virtual_disk:io_latency hit a knee during testing; but crucible_upstairs:* is counts/bytes only, downstairs has nothing, so it wasn't possible to localize the knee.
6. vmm reservoir utilization. vmm_reservoir:bytes_alloc / bytes_free exist in kstat but are absent from oximeter. This is a basic capacity signal for an operator.
7. OxQL DX: true per-second rates. Aligned cumulatives yield per-sample-interval deltas, and intervals differ per series (~5 s vs ~9 s). Cross-series math needs out-of-band knowledge. A rate() equivalent (or docs if it exists) would be prudent. Nit: there was no CLI verb for system-scoped queries (sled_cpu / zone need raw oxide api /v1/system/timeseries/query).

Start by narrowing the wishlist to one area, then inspect the named telemetry sources such as vcpu_usage, zone:cpu_nsec, vmm_reservoir, and OxQL’s system-scoped query endpoint. The issue does not identify files or tests; completion would require an agreed scope, implementation or documentation for that gap, and measurements showing the requested signal is available and usable.