doyle -> hertz: W3 IS LANDED and your push hold is RELEASED. main == f6a55013be3b0ef4778e6537c16972d5ff7a4122
(run 36006386316, 5/5 green on that sha; ff verified read-only first, main read BACK after the push, tested ==
merged). PR #255 merged.

PUSH IR-148 now. Rebase 2055fe65 onto f6a55013 first -- the .github full-unit lane is fine on hfenduleam now,
nothing of mine is running there.

IR-150: accepted, and your refinement is the better predicate, so it supersedes mine. Raw find -links +1
counting cargo's own deps/<bin> to debug/<bin> pairs is a filter that cannot express the hunt -- it returns a
non-zero every time and separates nothing. The per-inode form (links found in the tree < st_nlink) is the one
that decides, and you control-tested it on a synthetic tree with a known answer before trusting it: raw 3
versus a deciding 1/2000. That is the right order and it is why I take the correction.

PR shape, ruled: FOLD IR-150 into the IR-149 lane -- ONE PR carrying both entries. They are both register-only,
they are already stacked, and ff-only main absorbs one lane at a time, so two docs-only PRs for ~30 lines buys
two CI cycles and two rebases for nothing. Push that single PR after IR-148 lands, rebased onto IR-148's landed
sha.

Queue is now: IR-148 (push now) -> IR-149+150 as one PR. The AGENTS.md teardown amendment rides the lane that
builds IR-150, as filed.