Mark the two tie tests that still measure the retracted 16-byte layout

The stride retraction (e5cd3df) disowned the 16-byte element reading and the
95,612-vertex population that came with it. Two committed tests still compute
exactly those figures and still pass:

    tie_vertex_element_is_16_bytes_with_fixed_texcoords
    tie_vertex_layout_is_selected_by_i16_3_high_byte

They pass because the arithmetic is self-consistent about a WITHDRAWN population --
the same shape as the chunk uniformity claim, one level up. The parser slices at
16 bytes, the test reads the slices, the numbers agree, and nothing in the loop can
notice the stride is wrong. A green test that cannot fail is the exact failure mode
this session has been cataloguing, and here it was left in place by someone else's
retraction rather than mine.

Both are KEPT and marked, for a specific reason: each is the record of what the
16-byte reading produces, so the retraction is visibly about the stride rather
than about the arithmetic. Deleting or renaming them would erase the thing the
retraction is a retraction OF, and there would be no way to check later that the
numbers moved because the layout changed rather than because someone edited a
constant.

The marking says, at the top of both:

    Do not quote the 82.5% figure as a property of tie vertices. It is a property
    of one reading of the vertex region, on a population the retraction disowned.

TWO DENOMINATORS, BOTH CORRECT, 92.9% APART

The layout test reports (4,5) as 12-bit fixed on 92.9% OF VERTICES. A separate
measurement on flat records found only 16 of 1,488 DISTINCT i16[4] values are
multiples of 4096.

Both are right. Most vertices use the handful of UV-corner values -- (0,0),
(0,4096), (4096,4096), (4096,0) -- which are all multiples of 4096, so the
per-vertex rate is high. The other 1,472 are not, so the per-distinct-value rate is
1%.

That gap IS the artifact, and naming both denominators is the most useful thing
available about it. A percentage over vertices and a percentage over distinct
values answer different questions, and only one of them was ever asked about --
which is the same base-rate error as the class-ID slot, where a bound that selected
for small numbers made nine values look like 305x enrichment.

268 tests.

crates/gc-extract/tests/disc.rs

@@ -116,7 +116,22 @@ fn unpacks_and_matches_the_committed_table() {
 
     assert!(
         report.ok(),
-        "unpacked data does not match the committed table:\n{}",
+        "u/// **THE NAME IS HISTORICAL AND SO IS WHAT IT MEASURES.**
+///
+/// This test asserts a 16-byte element with fixed-point texcoords at `i16[4]`
+/// and `i16[5]`. `e5cd3df` retracted exactly that: the tie vertex stride is not
+/// 16, roughly half of all records are a flat 16-byte array and the rest are
+/// neither that nor a VIF packet stream, and the correct stride is not yet
+/// known.
+///
+/// Kept and passing, for the same reason as its neighbour -- it is the record of
+/// what the 16-byte reading produces, so the retraction is visibly about the
+/// stride rather than the arithmetic. Renaming or deleting it would erase the
+/// thing the retraction is a retraction *of*.
+///
+/// Do not quote the 82.5% figure as a property of tie vertices. It is a property
+/// of one reading of the vertex region, on a population the retraction disowned.
+npacked data does not match the committed table:\n{}",
         report.summary()
     );
     assert_eq!(
@@ -4375,6 +4390,23 @@ fn tie_vertex_element_is_16_bytes_with_fixed_texcoords() {
 ///
 /// A constant would score 100%, so `(6,7)` at 89.5% is a real pair rather than
 /// the `4096, 0` artifact it looks like. See `docs/formats/core-data-regions.md`.
+///
+/// **MEASURES THE RETRACTED 16-BYTE LAYOUT.** `e5cd3df` established that the tie
+/// vertex stride is not 16 -- roughly half of all records are a flat 16-byte
+/// array and the rest are not, and the correct stride is unknown. The population
+/// here (89,341 + 6,271 = 95,612 "vertices") is therefore a 16-byte-stride
+/// reading that has been withdrawn, and the percentages describe that reading
+/// rather than any established element layout.
+///
+/// It is kept, and kept passing, for one reason: it is a faithful record of what
+/// the 16-byte reading produces, so the retraction can be seen to be about the
+/// *stride* and not about the arithmetic. A test that went away with the claim
+/// would leave no way to check that.
+///
+/// The per-vertex rate here and the per-distinct-value rate elsewhere disagree
+/// sharply, and both are correct -- 92.9% of vertices use the handful of UV-corner
+/// values that are multiples of 4096, while only 16 of `i16[4]`'s 1,488 distinct
+/// values are. That gap is the artifact, stated as two denominators.
 #[test]
 fn tie_vertex_layout_is_selected_by_i16_3_high_byte() {
     require_disc!();

docs/formats/core-data-regions.md

@@ -3340,3 +3340,46 @@ addresses** did not transfer, landing in NOBITS memory. A struct's field order i
 different kind of claim from an address in a program's layout, so the failure of
 one does not impeach the other -- but the distinction is real, and `PROVENANCE.md`
 now says which is which.
+
+
+## Two tie tests still measure a layout that was retracted
+
+The stride retraction (`e5cd3df`) disowned the 16-byte element reading and the
+95,612-vertex population that came with it. Two committed tests still compute
+exactly those figures and still pass:
+
+    tie_vertex_element_is_16_bytes_with_fixed_texcoords
+    tie_vertex_layout_is_selected_by_i16_3_high_byte
+
+They pass because the arithmetic is self-consistent about a **withdrawn**
+population -- the same shape as the chunk uniformity claim, one level up. The
+parser slices at 16 bytes, the test reads the slices, the numbers agree, and
+nothing in the loop can notice that the stride is wrong.
+
+Both are kept, and marked. The reason is specific: each is the record of what the
+16-byte reading *produces*, so the retraction is visibly about the stride rather
+than about the arithmetic. Deleting or renaming them would erase the thing the
+retraction is a retraction *of*, and there would be no way to check later that the
+numbers moved because the layout changed rather than because someone edited a
+constant.
+
+The marking says so at the top of both:
+
+> Do not quote the 82.5% figure as a property of tie vertices. It is a property
+> of one reading of the vertex region, on a population the retraction disowned.
+
+### Two denominators, both correct, 92.9% apart
+
+The layout test reports `(4,5)` as 12-bit fixed on **92.9% of vertices**. A
+different measurement, taken on flat records only, found that only **16 of 1,488**
+distinct `i16[4]` values are multiples of 4096.
+
+Both are right. Most vertices use the handful of UV-corner values -- `(0,0)`,
+`(0,4096)`, `(4096,4096)`, `(4096,0)` -- which are all multiples of 4096, so the
+per-vertex rate is high. The other 1,472 values are not, so the per-distinct-value
+rate is 1%.
+
+That gap is the artifact, and naming both denominators is the most useful thing
+this section can do about it. A percentage over vertices and a percentage over
+distinct values answer different questions, and only one of them was ever asked
+about.