Automatic disk-pressure GC didn't trigger; volume reached 99% before manual GC

[schickling-assistant]

Observed

A macOS host running Determinate Nix reached 99% disk usage on its root volume before any automatic GC fired. Running nix-collect-garbage -d manually then freed 268.5 GiB across 26,023 store paths in a single pass on a ~927 GiB volume (used dropped from ~916 GiB to ~627 GiB; free climbed from ~11 GiB to ~300 GiB).

There were no GC-related entries in determinate-nixd logs (log show --predicate 'process == "determinate-nixd"' --last 7d) leading up to this.

Expected

Per the docs and prior discussion (e.g. #169), the managed GC strategy is "automatic" / disk-pressure-based, intended to keep ~5–20% free. On a ~927 GiB volume that would mean triggering somewhere around 46–185 GiB free; instead it sat at <2% free until manually invoked.

Environment

• Determinate Nix daemon + client: 3.17.1 (happy to retest on 3.18.1 if the strategy/threshold changed)
• macOS 26 (arm64)
• Single APFS volume, ~927 GiB
• /etc/determinate/config.json contains only a builder block; no GC overrides
• nix-darwin managed; no explicit nix.gc.* set

Questions

1. What is the actual trigger condition for the disk-pressure GC in the current implementation?
2. Is there a way for an operator to inspect the daemon's last GC decision / next-trigger threshold?
3. Is there a known failure mode where the watcher silently stops?