feat(grafana): add Kubernetes 3831 and 1471 curated packs

[shmsr]

Summary

• Curated pack for Grafana 3831 (Kubernetes Cluster Autoscaler): every panel is ES|QL because native PROMQL LAST(value, step) is empty in Lens on prometheus_native ingest. Grafana's bare sum() on *_total stays LAST_OVER_TIME (not RATE); time()-last_activity is DATE_DIFF seconds; Nodes available is ready/total * 100 so the number+% tile shows 100%, not 1%. Flattened "New row" panels sit at y: 2. KPI chrome titles are shortened for Kibana tiles.
• Curated pack for Grafana 1471 (Kubernetes App Metrics): Heapster container_name / pod_name / kubernetes_io_hostname and HTTP kubernetes_namespace rewrite to canonical labels. $container binds both cAdvisor container and HTTP app. Request rate restores nginx (engine dropped it for status vs code grouping) as Grafana-style native | 200 / nginx | 500 / haproxy | 2xx series. Plugin rewrites $namespace/$container populate off container_name. Per-pod graphs group by pod; response time uses PERCENTILE of the duration gauge.
• Engine: blank inner Lens labels (not just value / computed_value) so flattened metric tiles do not show the raw field name.
• Shared k8s_exporter now emits cluster_autoscaler_* plus cAdvisor spec/user/system and native/nginx/haproxy HTTP series so both packs validate on the curated rig.

Test plan

• pytest tests/test_curated_packs.py -k "3831 or 1471" (18 passed)
• Full tests/test_curated_packs.py + ruff + pin verifier
• Cross-profile leakage gate (scripts/run_cross_profile_corpus.py) — OK
• Live migrate + upload both dashboards (metrics-k8s.prometheus-default, prometheus_native)
• live_validate --fail-on-bug: 3831 10/10 ok, 1471 13/13 ok, 0 data_gap, 0 bugs
• dashboards_api --fail-on-error: 10 and 13 mapped, 0 errors
• grafana-validate-uploaded: 0 runtime errors, 0 empty, 0 overlaps
• Kibana view-mode: 3831 KPI values match exporter (2 nodes, 100% available, 0/1 safe, duration tiles); 1471 Request rate legend is native | 200 / nginx | 500 / haproxy | 2xx; $container=web → web-0/web-1, switch to redis → cache-0

Fixes #402
Fixes #403

[giorgi-imerlishvili-elastic]

Verdict: fixed, then approved. No Critical or Important findings remain.

What I verified

• Contract/diff: both canonical dashboards (Grafana 1471 rev 1 and 3831 rev 1) resolve through the new registry entries; both provenance pins verify. I generated and inspected IR plus native artifacts from the canonical sources: 23/23 panels mapped (1471: 10 clean + 3 explicitly approximate; 3831: 10 clean), 0 manual, 0 not-feasible, canonical namespace/container controls, single Lens breakdowns, guarded ratios, cumulative LAST_OVER_TIME semantics, DATE_DIFF activity tiles, and non-overlapping 3831 layout.
• Boundaries/regressions: reviewed empty/duplicate Lens labels, zero-denominator guards, optional HTTP-series behavior, canonical/legacy label rewriting, flattened-row positions, and the metric/gauge native payload shapes. The label change remains limited to empty or duplicate inner labels when chrome titles are shown. Five-profile portability passed without canonical-name leakage. No supported panel type or fidelity baseline was added/changed.
• Tests: the new targeted tests fail on the unmodified merge base at the first missing 3831 registry assertion, so they are not vacuous. The split 1471 pod-count test retains every original assertion.
• Docs/conventions: operator-visible behavior is documented in docs/sources/grafana.md; fidelity manifests contain no UNKNOWN entries and explicitly disclose the three 1471 approximations and three 3831 presentation approximations. The PERCENTILE concern was probed: it is not claimed as exact histogram_quantile parity; the manifest and emitted warning explicitly identify the duration-gauge approximation.

Fixes in b0c50e3

• Critical <parity-rig/curated/grafana_763_redis_exporter/k8s_exporter.py:257>: removed an unnecessary int() around an already-integer memory-limit expression. It mattered because Ruff RUF046 made the required CI gate fail. The emitted value remains byte-for-byte 536870912.
• Minor <tests/test_curated_packs.py:5714>: separated the Number-of-pods scenario from the response-time test so failures identify the behavior they cover.
• Minor <tests/test_curated_packs.py:5880>: removed the extra blank line at EOF so git diff --check is clean.

Gates

• Targeted: 19 passed, 204 deselected.
• Full unit suite: 6,219 passed, 53 skipped, 473 subtests passed.
• Lint: 498 source headers checked; skill mirror/structure checks passed; Ruff 0 errors.
• Typecheck: 0 issues in 10 source files.
• Provenance: 2/2 pins verified.
• Cross-profile corpus: 5 profiles passed; each produced 23 panels, 20 green / 3 yellow / 0 red.
• Post-push GitHub CI: 14/14 checks passed.

Visual verification did not apply to the review fix: it changes no migration/render/layout/control code. The only runtime edit is a lint-equivalent integer expression whose exporter output was verified identical; the other edits are tests/whitespace. No pre-existing base failures remain, and there are no optional Minor follow-ups from this review.