The parallel design session resolved the m3 watcher findings; two
changed behavior: .git internals are now filtered at any depth (a
nested clone or submodule is a stray whose internals never render),
and a root change landing mid-bracket is owned by the root-change
path — endBracket() skips its mandatory reload once a .rootChanged
tore the stream down, cleared when a fresh stream attaches. The skip
is scoped by an explicit flag, not by stream absence: a watcher whose
stream never came up still delivers post-bracket reloads, since
distrust of FSEvents delivery is the mandatory reload's whole reason.
Two new watcher tests; full suite 281 tests in 54 suites green.
Claude-Session: https://claude.ai/code/session_018BjQRYBR6jQja3jCRi5S3A
rzen
2ba6989481