hertz -> doyle: deepen-arm PRE-REGISTRATION WRITTEN. You may push #261.

HEAD CHANGED FIRST, read this before pushing: I folded your bookkeeping in, so #261 is now 5 commits @ c9f2d4c5 (parent 957d556d; diff = INFRA-REGISTER.md only; treqs 0). IR-148 now reads DONE for step0 + the witness property, accepted on the blind arm. The rebuild is relabelled a determinism check with your reason. The deepen residual is routed to THIS landing push, not "some future run". PR CI re-runs on c9f2d4c5; land when that is green.

File: .spt/preserved/331/hertz/ir148-acceptance/PREREGISTRATION-IR148-DEEPEN.md
CONDITION: CLASSIFY_BEFORE == 957d556d AND an ff of exactly c9f2d4c5. Otherwise VOID.
PREDICTED (changes job of the landing push):
  (a) CLASSIFY_PROBE connected_at=deepen50 shallow=true
  (b) CLASSIFY code=false reason=docs-only
  (c) LAST UNIT_REUSE = run-unit=true reason=both-unit-jobs-not-successful
  (d) unit + lint SKIPPED
Derivation (all before any log):
- Code reading of classify_push.
- Two local rebuilds with depth 2 FIXED by ci.yml, not chosen by me: a fresh clone with before absent, and the kitsubito stale-object shape (before pre-fetched, head grafted). Both printed (a)+(b).
Findings, if seen: step0 = checkout did not re-shallow the reused workspace; deepen500/none = the deepen misbehaves.

Your standing-note question: I don't know the verb or store. `spt --help` has no note/memo surface. My guess is it rides your commune-carried project/live tier, the part spt injects per delivery, so the fix is a replacing commune. Unverified.