refactor(chart): apply three dashboard conventions across every board

[Tanguille]

First of four. The three layers stacked on this branch (operator-health row, SGLang panels, llama.cpp board) follow one at a time as this merges, so each arrives as its own reviewable diff rather than one large change.

Adds no panels and no features: it exists so those layers inherit one convention instead of each inventing their own.

What

Convention	Applied to
min: 0 / max: 1 on percentunit ratios against a hard limit	llmkube-quota ids 1, 5
noValue: "0" where the outer PromQL is sum()/count()	model-router 1/2/4, sglang 1/6/8/14/15, vllm 1/6
palette-classic-by-name where series vary with a template variable	30 timeseries panels across sglang, vllm, llmkube-slo

Plus TestDashboardConventions in internal/metrics/dashboard_sync_test.go, so a future dashboard cannot silently violate any of the three. It walks fieldConfig.defaults per panel, pairing each panel with its own target exprs, and scopes the rules structurally rather than with an exemption list: error budgets (1 - (...)) and burn-rate thresholds (14 * (...)) fall out of rule 1 by shape, and sum(a)/sum(b) ratio-of-sums falls out of rule 2 because the sum() match does not close at the end of the string.

Why

palette-classic assigns colour by series order, so filtering $service repaints the survivors and a reader who learned "service X is green" is misled. Ratio panels without a max autoscale, so a 2% blip renders identically to saturation. count() over an empty vector reads "No data" on a fresh cluster instead of 0.

Not in scope

Legend calcs (["mean","max"] is the existing convention and stays), fillOpacity (a drawStyle: bars panel legitimately needs a high fill), colours on fixed series sets, and config/grafanas two boards, which predate this pass.

Evidence

• Audited all 7 boards, not only the 5 changed: amd-gpu-observability and llmkube-inference have no qualifying panels.
• Diffing both trees with min/max/noValue/color stripped is byte-identical on all 7 files, so no legend, fill, panel-id or formatting churn.
• helm unittest charts/llmkube 60/60, go test ./internal/metrics/... ok, all dashboard JSON valid.
• The test was falsified rather than assumed: removing one noValue makes it fail naming the exact panel, restoring makes it pass. It has already caught six real violations in the layers stacked on top, two of them in a commit that claimed to have fixed them.

Happy to split or reorder if you would rather see the test separately from the JSON changes.

Live validation

Chart rendered from all four branches merged, applied to a production cluster (AMD R9700, SGLang v0.5.16 + three llama.cpp services) behind a suspended HelmRelease, then reverted and proved clean.

33 of 37 panels return live data. The four empties: two are histogram_quantile over an idle operator, one was the queue-wait panel awaiting its recording rules (which materialised on apply at p95 18.0s / p50 1.0s), and one was a real defect in the llama.cpp branch, now fixed there.

Revert proof: recording rules back to the 5 the release ships, dashboards back to the 2 Flux owns, zero suspended Flux objects.

[Defilan]

Tanguille, this is one of the cleaner refactors I've seen, thank you. Zero PromQL touched (I diffed it: 0 changed expr lines), the key-change tally matches your description exactly, and the guard test is genuinely falsifiable, it already caught 41 real inconsistencies in the pre-PR dashboards. One thing before I approve: please add the AI-assistance disclosure line to the PR body per CONTRIBUTING (Assisted-by: <tool> (what it did; what you verified)). Add that and I'll approve and merge this one first, since #1369/#1370/#1371 all stack on it and will rebase onto it.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!