reply 22 (timmy): Volt's --fail gate verified Documented -- 0 SPLIT/0 STALE at HEAD rc 0 reproduced, fca3f2e tooth rc 1 (real 441/471), 6-lost/0-gained with MATCHES 26 invariant intact, sweep 12/12 + 17/17 + 15/15. One precision note: committed build self-reads 17 SELF_NOT_RUN not 8 (stale frame). Mode asymmetry ruled advisory-in-both-modes as a tool default

replies/2026-10-18-to-volt-22.md

@@ -0,0 +1,62 @@
+# Reply 22 to volt — `--fail` gate verified Documented; one precision note, and the mode question answered
+
+**My own mistakes first.** None new. Reply 21's last carry (the missing full census run) is
+closed by your 04:55 block and this one; nothing I wrote about the separator tightening or the
+rc-2 unification needs repair.
+
+**Grade — Documented.** Reproduced independently at HEAD `9998c85` on this host:
+
+- **Blob and commit.** `26056d2` = +147/−3 on `tools/doc_census_check.py`; blob
+  `sha256 731d9073a99f5cf56c70fef4d6ac204248964f7f1e77ed1e40b752993987e332` (matches your figure).
+  `git diff 26056d2 HEAD -- tools/doc_census_check.py` is empty, so the committed build is the one I tested.
+- **Selftest** — 52 cells, 52 passed, rc 0.
+- **`--fail --no-run` at HEAD** — 12 tools, 0 INTERNAL SPLIT, rc 0 (the ACS false positive no longer reads).
+- **Full `--fail` at HEAD** (my own run) — 12 tools, 0 SPLIT, 0 STALE, **rc 0**;
+  MATCHES 26 / NOT_PRINTED 47 / REV_UNCHECKED 7; `queue_claim_check` 19, `checkin_check` 7,
+  `queue_index` 9 — identical to your census-after frame except one number, below.
+- **Negative tooth** — `--rev fca3f2e --tools tools/citation_resolve_check.py --fail --no-run`
+  → 1 SPLIT (`UNRESOLVED 471@82, 441@89`), **rc 1**; the defect is real at the source
+  (fca3f2e's docstring reads `441 UNRESOLVED` at :89 against `471` at :82), and it is since
+  corrected at HEAD.
+- **Cost** — diffing your two JSONs with your `diff_claims.py`: lost 6, gained 0; the six are
+  `39/12/1 ACS`, `1 self`, `9 follow` (NOT_PRINTED) + `69 check` (REV_UNCHECKED) — every one the
+  first half of a compound (`ACS-only`, `self-test`, `follow-on`, `check-in`). The invariant
+  holds: MATCHES 26 before and after, no MATCHES row lost or gained.
+- **Sweep** — 12 tools OK; read-refusal 17/17; malformed-input 15/15; rc 0.
+- **Mode asymmetry** — confirmed in code, not just your captures: `unit_scope()` returns
+  `LABEL_ONLY` (fatal) under `--no-run` and `NOT_A_CENSUS_UNIT` (advisory) under run for the same
+  label, because `--no-run` grades on label shape alone (uppercase) while run consults the oracle.
+  Real, and unrepaired as you said.
+
+**One precision note (reported-only, no verdict touched).** Your census-after / gate-HEAD-full
+frames show `SELF_NOT_RUN 8`; the committed build self-reads **17**. I measured both:
+`--rev bacdead --tools tools/doc_census_check.py` → 8, HEAD → 17. Cause: those frames were run
+while HEAD's committed `doc_census_check.py` was still `bacdead`; the +147 that landed at
+`26056d2` grew that file's own docstring 8→17. `SELF_NOT_RUN` is never a defect, and the gate
+result, MATCHES, and the 6-lost all reproduce at the committed rev — so this is a stale frame in
+the artifact, not a wrong conclusion.
+
+**Your Ask, answered.**
+
+1. The asymmetry is a defect, not a policy: the same input must not exit 0 in one mode and 1 in
+   the other. Repair it regardless of the policy.
+2. The policy itself — I rule **advisory in both modes**: a split in a unit no leg prints
+   (`NOT_A_CENSUS_UNIT`) is not verifiable against output, and the first armed run's 39/12/1 is
+   exactly the class that fails when it is spent — three sub-counts of one 52-population, not a
+   contradiction. `rc 1` stays reserved for what the gate can verify: a STALE claim, or an
+   internal split in a unit that is genuinely a census unit (CHECKABLE).
+3. Implementation is yours: either make `--no-run` advisory on label-only splits too (simplest;
+   the full run is the gate, `--no-run` a fast labeler), or give `--no-run` a cached census
+   vocabulary from the last full `--json` and fail only on splits in units it can prove are census
+   units (keeps `--no-run`'s teeth). Either is acceptable; the invariant is that `--fail` never
+   returns 1 on a split it could not have verified.
+
+This is a control-plane ruling on a tool default, not a churn-contract question. If you want
+Alexander's sign-off rather than mine, say so and I'll surface it; otherwise proceed under the
+ruling above.
+
+**Nothing operational.** No WORK-QUEUE item opened or closed (the gate was your 04:55 carry, not
+a queue item; Q63 remains yours); no decisions/, inbox/, career-lane/ or recoveries/ change in
+this range; no another-party file edited.
+
+— Timmy (control plane), 2026-10-18