2026-05-16 — AKS card information density pass

Motivation

After the first v3 dashboard rollout the AKS card carried three problems the user surfaced explicitly during review:

1. The card body duplicated the per-node table that lives in the modal's Cluster Diagnostics section, but used raw m/Mi units that non-K8s operators can't read at a glance (257734Mi for an E32s_v5 is hard to recognise as ~252 GiB).
2. Three nodes in the same VMSS were rendered with identical truncated names because the replace(/-vmss\d+$/, '') rule stripped the only characters that disambiguated them.
3. Three "OK" signals (top-right OK chip + cluster-name Running chip + STATE Succeeded pill) competed for attention without conveying additional meaning in the steady-state.

Eight smaller items (#5..#11 in the analysis) were folded into the same pass to avoid a second round-trip.

User-facing change

Card body (the always-visible part)

• Replaced the per-node table with a one-line summary strip that reads:

  ● 1 (system) · ● 3 (user) · 4 NODES
   · CPU 0.52 / 98.0 cores (0.5%)
   · MEM 4.88 / 763 GiB (0.6%)
   · ● all Ready
  

• Pool dots colour-code system (warning/orange) vs user (accent/blue).
• Cores and GiB values are humanised (modal already had this; card now matches).
• Health flag on the right swaps to 1 NotReady (danger pill), MemoryPressure (warning pill), or 1 hot (warning) when the aggregate detects unhealthy/saturated nodes.
• The whole strip is clickable and opens the modal for the per-node breakdown — preserving the drill-down without rendering the rows twice on the dashboard.
• View full details button kept for keyboard / screen-reader access.

Cluster header row

• Workspace ready chip moved from its own standalone strip into the same row as the cluster name and power state. The "is this cluster usable?" answer now lives in one place.
• Stop / Delete buttons are now grouped at the far right of the row, separated from the cluster name by a thin vertical divider, so destructive actions don't sit shoulder-to-shoulder with metadata.
• The standalone "Workspace ready" strip is removed; the per-database warmup chip strip remains so progress is still visible while DBs are loading.

STATE pill

• Hidden when ARM provisioning is the steady Succeeded state — the Running chip already conveys "the cluster is fine". Still rendered (with the appropriate accent / warning / danger pill) for Creating, Updating, Deleting, and Failed.

SHARD CAPACITY block

• Rewritten as a single line: Shard capacity · 3 nodes · Standard_E32s_v5 · auto-sized per submit (max 10 jobs).
• target N ≤ 10 jargon replaced with max 10 jobs; the sentence about ARM throttling lives in the explanatory title tooltip.

Pool cards

• Each pool card now carries a footer line with a per-node and total capacity readout, e.g.:

  SYSTEM   1 × Standard_D2s_v3
  2 cores · 8 GiB / node
  

  USER     3 × Standard_E32s_v5
  32 cores · 256 GiB / node · 96 / 768 GiB total
  

  Total only renders when the pool has more than one node (otherwise it would just repeat the per-node value).

API / IaC diff summary

Frontend only — no backend or Bicep changes.

File	Change
web/src/components/ClusterDetailModal.tsx	ClusterDetails body table replaced with a memoised aggregate summary strip. Helpers isSystemPool, fmtCores, fmtGiB colocated. Imports trimmed (Maximize2, AlertTriangle, Copy, Loader2, X only).
web/src/components/ClusterItem.tsx	useAksSkus consumed for pool capacity totals. Workspace ready chip moved next to power label. Stop/Delete grouped on the right with a divider. STATE pill hidden when Succeeded. Shard-capacity block reflowed to one line with friendlier wording.

No new dependencies, no CSS changes (existing dv3-pool-card .footer, dv3-warmup-chip, dv3-shard-capacity tokens are reused).

Validation evidence

• npx tsc --noEmit (in web/) → clean.
• npm run build (in web/) → success (671 kB JS / 89 kB CSS, unchanged chunk warning).
• uv run pytest -q api/tests → not re-run; backend untouched in this pass (the previous change note in this directory ran 219/219 already).
• Visual verification at http://127.0.0.1:18080/:
• Card body renders the new one-line summary ● 1 ● 3 4 NODES · CPU 0.52 / 98.0 cores (0.5%) · MEM 4.88 / 763 GiB (0.6%) · all Ready.
• Running 🔥 ready chips sit next to the cluster name; Stop/Delete are right-aligned behind a divider.
• No redundant STATE Succeeded pill.
• Pool cards show 32 cores · 256 GiB / node · 96 / 768 GiB total.

Out of scope

• #9 — cascade pressure to the cluster card header dot is partially addressed by the NotReady / pressure / hot pill in the new summary strip. A full propagation up to the cluster name row would require hoisting the topQuery from ClusterDetails into ClusterItem. We decided that the prominent in-strip pill is enough for now; revisit if a real incident shows operators missed it.
• #8 — cross-pool memory % comparability hint is moot at the card level now that the body shows a single cluster-wide aggregate. The per-pool grouping inside the modal already distinguishes system vs user nodes by colour stripe.