fix: normalize CPU % against container cpuLimit and add soft ceiling to BarChart

[amir20]

Summary

• CPU normalization fix: When a container has a CPU limit (--cpus=N), the frontend was dividing raw CPU% by 1 instead of N. A --cpus=2 container using 1.5 cores now correctly shows 75% instead of being capped at 100%. Fixed in ContainerStatCell, HostCard, and MultiContainerStat.
• BarChart soft ceiling: Added maxCeiling prop (default 10) so bar scaling uses max(dataMax * 1.2, maxCeiling) instead of just max(data). Idle containers stay visually flat instead of filling the full chart height. Spikes are visible relative to recent history.

Test plan

• Run docker run -d --cpus=2 alpine sh -c "for i in 1 2 3 4; do while true; do :; done & done; wait" and verify Dozzle shows ~100% (not capped at 100% from raw 200%)
• Run a container with no CPU limit and verify it still normalizes against host CPU count
• Verify idle containers show small bars in the container table, not full-height
• Verify a CPU spike shows meaningful visual variation in the bar chart
• Run pnpm test and pnpm typecheck — both pass

🤖 Generated with Claude Code

[claude]

Review

CPU normalization fix - Correct. calculateCPUPercentUnix multiplies by onlineCPUs, so a container with --cpus=2 at full utilization reports 200% raw. The old return 1 divided by 1, overcapping at 100%. Dividing by cpuLimit (2.0) gives correct 100%. Fix is consistent across all three files.

BarChart soft ceiling - Improvement for relative scaling, but one behavioral note:

• MetricCard in HostCard.vue passes totalCPU which is a sum across all containers (can exceed 100%). Old code capped at 100% height; new code scales relatively — actually better, but data passed as percent to MetricDataPoint semantically implies 0-100 and could exceed that on multi-container hosts. Not a blocker.
• StatMonitor.vue passes raw y-values (not percentages). Old Math.min(dataPoint, 100)% would incorrectly cap non-percentage data. New relative scaling is strictly better here.
• The maxCeiling = 10 creates a visual discontinuity around 8.3% (where dataMax * 1.2 overtakes the ceiling). Containers hovering near that threshold will see the scale shift. Minor visual edge case.

No bugs, no security concerns. The CPU normalization is the substantive fix here.