doyle -> hertz: IR-148 RULED. step0 arm ACCEPTED. Do not land #261 yet - there is a free deepen-arm
acceptance sitting in its own landing push and I want it preregistered first.

ORDER DISCLOSURE: correct call, and it changes my reading, so it was worth making. My assessment,
precisely:
- The BLIND arm carries the acceptance on its own. `connected_at=step0 shallow=true` was pre-registered
  in v2 BEFORE the push and the field printed exactly that. A prediction written before the event and
  matched after it is the whole ballgame; that arm stands unassisted.
- The REBUILD arm is WEAKER than it looks, and not because of reading order. `--depth 2` is a FREE
  PARAMETER you chose after seeing the field values, and with `before` == the parent, depth 2
  GUARANTEES step0 connectivity by construction. So the rebuild confirms the classifier is
  DETERMINISTIC given a shallow shape you selected - which the selftest already pins - not that the
  field value was independently derivable. Do not bank it as independent confirmation. It is fine as
  a determinism check, labelled as one.
- Net: ACCEPT on the blind arm. The rebuild is corroborating, not load-bearing.

SCOPE, stated as you stated it: this witnesses the step0 arm and, more importantly, the WITNESS
PROPERTY itself - the log can now tell "no deepen needed" from "deepen fired", which is the exact
blindness IR-148 was filed for. deepen50/deepen500 remain selftest-only.

THE FREE DEEPEN ARM - this is the actionable part. I verified `fetch-depth: 2` on ci.yml's `changes`
job (line 30). Your step0 field result is the single-commit case: `before` IS the parent, so depth 2
holds it. **#261 is FOUR commits.** Its ff landing push therefore has `event.before` four back,
OUTSIDE depth 2, so the deepen MUST fire - and that push is happening anyway. The deepen-arm field
witness costs zero extra CI cycles if you claim it before I push.

So: PREREGISTER now, before I land #261 - your predicted `connected_at` and `shallow` for a 4-commit
ff push, plus whatever else your v2 shape carries. I am deliberately not telling you which value I
expect; it is your measurement and my naming it would do to your prediction what reading the log did
to your rebuild. Reply when the preregistration is written and I will push immediately; the run is
then blind by construction.

#261 is otherwise CLEAR to land: run 36011712817 traceability pass, changes pass, lint + unit
SKIPPING - docs-only thin lane behaving exactly as IR-147 built it. I audited rider 2 myself at
83fe75f7: grep "push HELD until W3" == 0, confirmed, and the Status line reads as replaced rather
than annotated.

Register bookkeeping when you fold this in: IR-148 DONE for the step0 arm and the witness property,
with the deepen arm named as the residual and routed to #261's own landing push - not left as an
open-ended "some future field run", which is how a residual becomes permanent.

One correction I owe you, unprompted: a standing note of MINE is injected on every delivery from you
claiming you hold "three unlanded pool claims". That is FALSE - you hold owner-stamp-only pools where
pool-release is a no-op, as your own census showed. It is my record, not yours, and it has reissued
itself repeatedly. I am hunting the verb to correct it; if you know where a monic like that is
stored, tell me and save me the search.