docs: R-306 — the merged design reproduces the answer from two implementations

A second implementation of the three settled rules, written from the
description alone rather than from the peer's code or the agreed list, lands on
the same nine columns and the same single exclusion. Two unlike
implementations of the same three sentences agreeing is the property a design
needs before anybody builds it.

Also tabulates where each of the evening's counts came from: 44% and 57% and
64% and 'at least 7' and 8 each came from an operation on the numbers; 9 came
from opening the disputed columns and reading them.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

docs/REVIEW_LOG.md

@@ -7601,3 +7601,54 @@ answer.
 
 **Nothing in the design turns on it.** Seven, eight or nine, bold is unreliable inside tables,
 and the four-row reading of the mixed-force table settles the merge on its own.
+
+---
+
+## R-306 — the merged design reproduces the answer from two independent implementations
+
+**The argument for the fold is no longer a preference.** The concurrent session implemented the
+three settled rules and reported that they produce the agreed nine. Rather than accept that,
+this is a **second implementation written from the description alone** — not from their code,
+not from the agreed list — and it lands on the same answer:
+
+```
+union vocabulary + exclusion tested on cells + bold ignored as a signal
+
+  1119 "Wiped"  n=9 bold=4        1234 "Then wipes"          n=2 bold=1
+  1119 "Median rounds"  n=9 bold=4   1301 "Median"            n=2 bold=1
+  1143 "Wiped"  n=4 bold=2        1301 "Past 15 rounds"       n=2 bold=1
+  1143 "Median rounds"  n=4 bold=2   1301 "1 in 100 runs past" n=2 bold=1
+  1198 "Wiped"  n=5 bold=2
+
+genuine 9, excluded 1
+  EXCLUDED (cells are sentences)  THROUGH_TRAIN.md "Measured over 300 runs"
+```
+
+**Two implementations, written separately from the same three sentences, agree on all nine
+columns and on the single exclusion.** The rules are therefore specified well enough to build
+from — which is the property a design needs before anybody writes the merged guard, and it is
+the one thing this long exchange produced that neither session set out to get.
+
+**The exclusion is the part that carries it.** *"A column whose cells are sentences rather than
+values"* was implemented here by removing the figures and asking what English is left; the
+other session tested it differently and reached the same verdict. A rule that survives two
+unlike implementations is a rule, not a heuristic.
+
+### Where the counts came from, in order
+
+| Claim | Source of the number |
+|---|---|
+| 44% | one session's vocabulary, quoted as the other's |
+| 57% | the other's vocabulary, transcribed by hand |
+| 64% | the other's vocabulary, run by its owner |
+| "at least 7" | the intersection — an operation on the sets, not a look at them |
+| 8 | the union agreed, one session's own set counted |
+| **9** | **every disputed column opened and read** |
+
+**Every wrong number came from an operation. The right one came from looking.** That is the
+session's finding, and it held at four different scales: a vocabulary mistaken for a
+measurement, a column classified by its neighbours, an intersection standing in for an
+inspection, and seven columns asserted genuine inside the entry correcting exactly that.
+
+**Nothing merged.** Both guards remain in the chain, both green, and the decision is still with
+the humans — but it is now a decision about whether to build, not about what to build.