docs(agents): a bead may not close on a gate whose compilation was not observed

Encodes the rule that would have caught all four of the false-state closures
found this week: bd-9z5ikz, bd-8f6gck, bd-t5d6h2 and bd-r71n1b each closed on a
gate that could not link, so the gate reported nothing while the fix was
reverted underneath it. An unlinkable target is silent, not green.

Also records that 'cargo check --all-targets' STOPS at the first failing target
(2 reported vs 62 actual with --keep-going), that the Jun-26 deletion incident
has THREE commits not two (517d0a2 at -6914 lines is larger than the other
two combined), and that this class must be audited by per-commit deletion size
rather than by reading commit messages, since all three messages are accurate
about what they added.

Beads: bd-1m5mbs (fromfp reversion, fixed 72ae1c2), bd-om1s58 (posix_fallocate
lost its EOPNOTSUPP emulation, LIVE), bd-t5d6h2 reopened, bd-8f6gck annotated.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: Dicklesworthstone