fix(skills): correct perf tool-arg names + metric/runtime inaccuracies (code-review followup)

[rlaope]

Summary

Applies the verified findings from an extra-high-effort multi-angle /code-review of the 11 skill playbooks merged in #94. All are doc-level corrections (no Go change); the most important class — wrong perf.* argument names — was confirmed by reading the actual tool schemas in internal/core/tools/perf/.

CRITICAL — perf tool calls used non-existent arg keys (silently dropped → wrong endpoint/default window)

• node-runtime: v8_inspector_targets/cpu_profile used url= → real args are name / target_index / duration_seconds / top_n.
• go-runtime: go_pprof_cpu used url=/seconds= → real args name / duration_seconds / top_n.
• native-perf: linux_perf_record used duration= → real arg duration_seconds.

HIGH — PromQL / metric correctness

• capacity-scheduling: kube_pod_container_resource_requests / kube_node_status_allocatable summed per node without a {resource="cpu"|"memory"} matcher mixes cores + bytes → added the filter.
• slo-burn: time_to_exhaustion omitted the SLO window W (dimensionless result) → now W × remaining / burn_rate.
• ai-inference: vllm:gpu_cache_usage_perc is a fraction in [0,1] despite the _perc suffix → thresholds go against 1.0, not 100.
• go-runtime: mark-assist is concurrent inline work, not an STW pause (was "STW assist pauses"); throttle test now uses cfs_throttled_periods/cfs_periods, not a bare rate.

MEDIUM — runtime facts & routing

• node-runtime: dropped the bare node trigger (hijacked k8s cluster-node queries like "node NotReady"); nodejs_active_* are gauges (non-recovering rise, not monotonic counter); --max-old-space-size default is memory-derived on Node ≥ 12, not a fixed ~1.5 GB.
• dotnet-runtime: LOH threshold is 85,000 bytes (~83 KB), not "85 KB".

LOW — read-only guard / output-shape consistency

• slo-burn: added a dedicated "never silence alerts / edit rules" read-only bullet; gated the fixed output shape on T/W being known rather than invented.
• capacity-scheduling: marked the Recommend field as operator-applied/read-only (house style).

Test plan

• go test ./internal/core/skills/... — all 24 skills parse/load
• go test ./... — full suite green
• grep confirms no remaining url=/seconds=/duration=<n> perf args in any skill body
• perf arg names cross-checked against internal/core/tools/perf/{pprof_cpu,linux_perf,rbspy,v8,v8_cdp}.go