docs: R-305 — the union, not the intersection: nine columns

R-304's rate was an artefact; the proposed correction, retreating to the seven
both vocabularies agree on, overshoots. Opening every disputed column instead
of comparing totals: 'Then wipes' is genuine and check-unmarked misses it
because wiped? does not match wipes; '1 in 100 runs past' is genuine and
check-figures misses it because its vocabulary has rounds? and not runs; only
THROUGH_TRAIN's 'Measured over 300 runs' is a false positive, and there the
cells are whole sentences and the bold wraps a clause.

So nine, and the shape matters more than the number: each vocabulary has a real
gap the other covers, which argues for the union of both word lists.

Refuses one recommendation. Dropping runs? from the header vocabulary would
also drop the p99 column, whose cells cite fight-tail cut.p99 and column.p99
and which R-269 added because the median is not a plan. That is the R-299
failure again — a guard narrowing itself until the figure it exists to watch
falls outside. The real exclusion wanted is 'a column whose cells are sentences
rather than values', which is testable; subtracting a word cannot tell the two
cases apart.

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

docs/REVIEW_LOG.md

@@ -7516,3 +7516,57 @@ two symptoms.
 
 **No merge performed, and neither session will start one on the other's say-so.** The decision
 is with the humans; this entry is the evidence for it.
+
+---
+
+## R-305 — the union, not the intersection: nine columns, and one recommendation to refuse
+
+**R-304's rate was an artefact and the correction to it overshot.** The concurrent session
+proposed retreating to the **intersection** — *"at least seven under both definitions, and
+everything beyond seven is a property of the measuring instrument"*. Seven is right as a
+floor and wrong as the answer, because **the two extras are not symmetrical noise: one of
+each pair is a real column the other vocabulary cannot see.** Adjudicated by opening every
+disputed column rather than by comparing totals.
+
+```
+shared by both vocabularies   7   all genuine
+only check-figures            1   CLEAN_GROUND.md:1234  "Then wipes"
+only check-unmarked           2   CLEAN_GROUND.md:1301  "1 in 100 runs past"
+                                  THROUGH_TRAIN.md:709  "Measured over 300 runs"
+```
+
+| Disputed column | Cells | Verdict |
+|---|---|---|
+| **"Then wipes"** | `**25.5%**` and `7.8%` | **genuine.** `check-unmarked` misses it because `wiped?` does not match *wipes* |
+| **"1 in 100 runs past"** | `32.3` and `**45.3**` rounds, citing `fight-tail cut.p99` / `column.p99` | **genuine.** `check-figures` misses it because its vocabulary has `rounds?` and not `runs` |
+| **"Measured over 300 runs"** | whole sentences — *"**2 of 4 hurt, half a party member down, no wipes.** The intended shape."* | **false positive.** The bold wraps a clause, not a figure |
+
+**So the honest count is nine**, the union minus one verified false positive — and the useful
+finding is not the number but its shape: **each vocabulary has a real gap the other covers.**
+Mine cannot see *runs*, theirs cannot see *wipes*. That is a straightforward argument for the
+merged guard taking the union of both word lists, and it is better evidence than either total.
+
+⚠ **And one recommendation from that exchange must be refused.** It was proposed that the
+merged guard **drop `runs?` from the header vocabulary**, on the grounds that it catches
+thresholds and sample sizes. It would also drop *"1 in 100 runs past"*, whose two cells are
+`fight-tail`'s p99 figures and are cited as such. **The p99 column is the most load-bearing
+column in that table** — R-269 put it there precisely because the median is not a plan — and a
+vocabulary change that silently stops holding it would be the R-299 failure again: a guard
+narrowing itself until the figure it exists to watch falls outside.
+
+**The actual defect in the THROUGH_TRAIN case is not the word `runs`. It is that the column's
+cells are prose.** *"Untouched. A legitimate choice for a pass."* is not a figure, and bold
+there is sentence emphasis. The exclusion a merged guard wants is **"a column whose cells are
+sentences rather than values"**, which is a property of the cells and can be tested, rather
+than a subtraction from the word list, which cannot distinguish the two cases at all.
+
+**The design conclusion from R-304 is untouched**, and nothing here disturbs it: prose takes
+bold-as-published, tables take column-inherits-header with bold ignored entirely. That rested
+on the four-row reading of the mixed-force table, which needs no count whatever.
+
+**The meta-lesson, which is the one worth keeping.** Two sessions measured a property of the
+corpus with an instrument whose calibration was the very thing under dispute — the vocabulary
+was both the subject and the ruler — and produced 44%, 57% and 64% in turn, none of which was
+about the documents. The recovery is not to retreat to what both instruments agree on. **It is
+to open the disputed cases and look**, which is what distinguishes a real gap from noise, and
+which no amount of comparing totals can do.