diff --git a/.claude/board/LATEST_STATE.md b/.claude/board/LATEST_STATE.md
index 2bab24ee9..efd6cba4a 100644
--- a/.claude/board/LATEST_STATE.md
+++ b/.claude/board/LATEST_STATE.md
@@ -1,9 +1,14 @@
 ## 2026-09-16 — PR #1241 merged (`d5d3f7ab`): the facet's per-axis LCP is one masked register readout — NO contract inventory delta
 
-`FacetCascade::hi_distance` / `lo_distance` now read `tz((a ^ b) & AXIS_BYTES)
-/ 16` off the `u128` instead of gathering a six-byte chain and walking it:
-12.5 → 5.8 ns for both axes, bit-identical over 64K pairs, disable-verified.
-No public type added or removed; two private consts + one private `const fn`.
+`FacetCascade::hi_distance` / `lo_distance` now read
+`(tz((a ^ b) & AXIS_BYTES) − 32) / 16` off the `u128` — the 32 classid bits
+sit below the tiers, so the subtraction is load-bearing: without it a tier-0
+divergence reads as prefix 2, not 0 (Codex on #1242 caught the formula as
+first written here, which omitted it; the merged code always had it) —
+instead of gathering a six-byte chain and walking it: 12.5 → 5.8 ns for both
+axes, bit-identical over 64K pairs, disable-verified. No public type added or
+removed; two private consts + two private `const fn`s (`tier_byte_mask`,
+`shared_axis` — "one" as first written here was wrong, same review).
 Arc entry in `PR_ARC_INVENTORY.md`; the fold-once epiphany
 (`E-FORMAT-SLOT-FOLD-IS-THE-SAME-OP-AS-THE-VL-DESCENT-1`) landed with the PR.
 Sibling measurements the same day: ndarray #311 (ternlogq tail descent 5–8×;
diff --git a/.claude/board/PR_ARC_INVENTORY.md b/.claude/board/PR_ARC_INVENTORY.md
index b82a2c035..29346e9c1 100644
--- a/.claude/board/PR_ARC_INVENTORY.md
+++ b/.claude/board/PR_ARC_INVENTORY.md
@@ -9,6 +9,15 @@
   identical case. Board: `E-FORMAT-SLOT-FOLD-IS-THE-SAME-OP-AS-THE-VL-DESCENT-1`.
   2 files, 3 commits (the first cut re-folded the gathered chain into a
   `u64` — 1.5×; the operator's *"fold the PowerShell logic once"* replaced it).
+  - *Correction 2026-09-16 (Codex on #1242):* **4 commits**, not 3 —
+    `95e2863` (u64 fold), `ce6664e` (epiphany entry), `76bb237` (masked
+    single-register readout), `45c5199` (entry corrected to the readout).
+  - *Correction 2026-09-16 (Codex on #1243):* the **Locked** line below
+    reads `tz((a ^ b) & AXIS_BYTES) / 16` — it omits the `− 32` for the four
+    classid bytes below the tiers. The merged `shared_axis` is
+    `(tz((a ^ b) & AXIS_BYTES) − 32) / 16`; without the offset a tier-0
+    divergence would read as prefix 2, not 0. The line itself stays as
+    written (append-only); this note is the record.
 - **Locked:** *the register is already the formatted form; never un-format it
   to compute on it.* An axis prefix is `tz((a ^ b) & AXIS_BYTES) / 16` past
   the classid — the same op as the whole-facet `prefix_distance`.