docs: R-303 — the two guards are disjoint in both directions

R-302 showed check-unmarked catching what check-figures misses. I had taken
that, plus a preference for structure over vocabulary, as grounds for folding
check-figures in as a secondary pass. The converse test contradicts it: five
unbolded cells in a measurement column, markers stripped, fail check-figures
and pass check-unmarked, whose strict class requires bold by design.

So the relationship is symmetric. The shapes are the weak half and should fold
in behind the structural classes; the column-header rule is not a shape and
belongs beside bold-as-published, not under it.

Also records what a merged guard must face rather than inherit: this corpus
bolds inconsistently inside tables, which is invisible to a reader and
load-bearing for a class keyed on boldness.

Both runs in throwaway worktrees, never the shared tree. No merge performed;
the decision sits with the humans.

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

docs/REVIEW_LOG.md

@@ -7399,3 +7399,57 @@ had the file on disk.
 finding before accepting it, repaired it in `e9f7d45`, then checked my account of their own
 commit and found it wrong. R-300's technical claim stood; its account of how the breakage
 happened did not, and I had asserted a mechanism I never looked at.
+
+---
+
+## R-303 — the two guards are disjoint in both directions
+
+**Evidence for a decision neither session should make alone**, and it corrects a
+recommendation I made on one-way data. R-302 showed `check-unmarked` catching what
+`check-figures` misses; I had concluded from that, plus a general preference for structure
+over vocabulary, that mine should fold in as a **secondary pass**. **The converse test says
+otherwise.** Both runs in throwaway worktrees off `e2a474b`, never in the shared tree.
+
+**Their direction** — strip the markers from two bold prose figures, `**1.29**` and
+`**48.7%**`, neither in a table:
+
+```
+check-figures:  OK                                            exit 0
+check-unmarked: FAILED — 2 published figures carry no marker  exit 1
+```
+
+**The converse** — strip the markers from five *unbolded* cells in one row of the
+mixed-force table, every one of them in a measurement column:
+
+```
+| Six hollow men, column stands off | 3.48 of 6 | 0.42 | 0% | 0.04 | 13 |
+
+check-figures:  FAILED — 5 figures carry no citation marker   exit 1
+check-unmarked: OK                                            exit 0
+```
+
+| Defect | `check-figures` | `check-unmarked` |
+|---|---|---|
+| Bold figure in prose, marker stripped | misses | **catches** |
+| Unbolded cells in a measurement column | **catches** | misses |
+
+**So the relationship is symmetric, not hierarchical.** Bold-as-published owns prose and
+cannot see an unbolded cell; column-inherits-header owns tables and cannot see a bold
+sentence. `check-unmarked`'s strict class requires bold by design — those five cells fall into
+the twenty it counts loosely and leaves unjudged, which is correct for it and still means five
+stripped citations pass the build.
+
+**What was actually wrong in my recommendation.** Not the preference for structure over
+vocabulary — that stands, and my eight **shapes** are the weak half that should fold in behind
+`check-unmarked`'s classes. But **the column-header rule is not a shape.** It reads a claim the
+document makes about its own columns, and it belongs beside bold-as-published rather than
+under it.
+
+⚠ **And a fact a merged guard will have to face rather than inherit.** The two mechanisms
+disagree about what this corpus's tables are, because **the corpus bolds inconsistently inside
+them** — the same measurement is bold in one row and plain in the next, four lines apart. That
+is invisible to a reader and load-bearing for a strict class keyed on boldness. Better said
+out loud by whatever replaces these two than quietly depended on.
+
+**No merge performed.** Both guards remain in the chain, both green, and the decision sits
+with the humans.