[rocketraman]

Two follow-ups from the 1.6.1 diagnostics report on Apple Silicon.

Don't rescan for writes in git-ignored folders

In the report, one repo's build/ was behind 97% of its file events (28,488 events) and caused 224 rescans, about half of all the rescans the app made. While a build runs, the debounce ceiling rescans the repo every 5s, and git sees none of those changes.

• WatchRoots.ignoredDirs asks git which top-level folders are ignored (check-ignore). It leaves out any folder that still has tracked files in it (ls-files), and never asks about .git.
• WatchPolicy.actionFor drops an event only when all of its paths are inside one of those folders. These still trigger a rescan:
  • the folder itself being created or deleted;
  • files at the top of the repo;
  • a move from build/ into the code;
  • the watcher's lost-events signal.
• The answer is fetched once when a repo is registered. It is fetched again in the background when an event lands in a top-level folder that wasn't there last time, or when .gitignore or .git/info/exclude changes. Until it refreshes, nothing extra is dropped, so a stale answer costs a rescan and never hides a change.
• The folders are still watched. Not watching them at all stays blocked on Nucleus#571.
• Not covered: ignored folders below the top level (app/build/), and changes to the global excludes file before the repo is next registered.

Say allocation is unavailable when every counter reads 0

The native image reports per-thread allocation counters as supported, then reads 0 for every thread. The report showed allocated 0MB since start next to 218 young collections. It now reads alloc unavailable (this runtime's per-thread counters all read 0).

Testing

• 12 new tests:
  • git-backed: ignored vs tracked-in-ignored folders, .git, a folder that isn't a repo;
  • pure policy: drop vs rescan edges, and when the answer goes stale.
• Unit tests 289/289, Konture 10/10, detekt clean.
• Not yet run in the app against a live build.