What changed, and when

Run a fabric check on any cluster to see speed and retries for every link. The same check now runs during every launch.

· New command: ferrite fabric check --per-link
· Degraded links are flagged before a cluster is handed over
· Results are kept for 30 days in the console

· Fixed: monthly totals rounded twice when shown in EUR or GBP
· Fixed: invoice PDFs missing the VAT number on some accounts
· Fixed: usage chart skipping the last day of the month

Reserved blocks can now move up or down by up to 25% each month and keep the committed discount.

· Set a monthly range when you create a block
· Changes apply from the next billing day
· Usage above the block is billed on demand, by the second

Hall caches now hold the 200 most used training images in each region, up from 50. Most launches skip the registry pull entirely.

· Median launch time down from 112 to 84 seconds
· Pin your own images to the cache from the console
· Cache hits shown in the launch log

· Fixed: final checkpoint lost when a spot node was reclaimed mid-upload
· Fixed: ferrite run exiting with code 0 after a node failure
· Fixed: log streaming stopping after 24 hours

Forge 80 clusters are now available in Tokyo, with object storage in the same hall.

· Up to 256 Forge 80 GPUs per cluster
· Round trip from Singapore around 70 ms
· Same prices as our European regions

Alert rules now group related events, so one failing node sends one message instead of forty.

· Alerts grouped by cluster and node
· Email, chat and webhook targets
· On-call paged within 30 seconds of a hardware fault