re: n206 was fixed after all -- the audit was classifying by geometry, not buffer

debug_grouped_report (why_rejected) shows n206_02's alternative pool is ACCEPTED
at pad 0 under production gates and is in the candidate list -- and the decoder
does take it: n206_02 now anchors at 0x342d984. The 'still collapsed' reading
came from the audit comparing decoded geometry, and that offset holds a direct
(unmirrored) copy, so a separated pair still looked identical. The audit now
requires a SHARED BUFFER to call it a collapse: disc-wide 18 exact mirrors, 16
related, 0 collapses, 0 unrelated -- and the regression test drops its exception.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

crates/sylpheed-formats/examples/starts.rs

@@ -0,0 +1,13 @@
+use sylpheed_formats::mesh::{debug_resource_params, debug_vertex_run_starts};
+fn main() {
+    let a: Vec<String> = std::env::args().collect();
+    let bytes = std::fs::read(&a[1]).unwrap();
+    let (markers, stride) = debug_resource_params(&bytes, &a[2]).expect("resource");
+    println!("{}: stride={stride} markers={markers:?}", a[2]);
+    let starts = debug_vertex_run_starts(&bytes, stride);
+    println!("{} candidate starts at stride {stride}", starts.len());
+    for off in &a[3..] {
+        let o = usize::from_str_radix(off.trim_start_matches("0x"), 16).unwrap();
+        println!("  0x{o:x} in candidate list: {}", starts.binary_search(&o).is_ok());
+    }
+}

crates/sylpheed-formats/examples/twin_mirror_audit.rs

@@ -74,6 +74,15 @@ fn main() {
                 related += 1;
                 continue;
             }
+            // Identical GEOMETRY is only a collapse if it also comes from the
+            // same buffer: a container may store a part twice unmirrored, and
+            // then two twins on two copies decode identically and correctly.
+            let shared_buffer = m.meshes[0].vbuf_offset.is_some()
+                && m.meshes[0].vbuf_offset == t.meshes[0].vbuf_offset;
+            if ident && !shared_buffer {
+                related += 1;
+                continue;
+            }
             if ident {
                 same += 1;
                 if examples.len() < 6 {
@@ -101,7 +110,7 @@ fn main() {
     }
     println!("twin pairs of equal vertex count: {}", same + mirrored + unrelated + related);
     println!("  exact X-mirror (expected) : {mirrored}");
-    println!("  IDENTICAL  (collapse)     : {same}");
+    println!("  IDENTICAL, one buffer (collapse) : {same}");
     println!("  related other way (Y/Z mirror, reordered): {related}");
     println!("  unrelated  (mis-anchor?)  : {unrelated}");
     for e in examples {

crates/sylpheed-formats/examples/why_rejected.rs

@@ -0,0 +1,14 @@
+//! Why does the decoder refuse a grouped-pool resource at a given pool start?
+//! Usage: why_rejected <container.xpr> <resource>@<vb0-hex>...
+use sylpheed_formats::mesh::debug_grouped_report;
+fn main() {
+    let a: Vec<String> = std::env::args().collect();
+    let bytes = std::fs::read(&a[1]).expect("container");
+    for pair in &a[2..] {
+        let (name, off) = pair.split_once('@').expect("name@hex");
+        let vb0 = usize::from_str_radix(off.trim_start_matches("0x"), 16).expect("hex");
+        for line in debug_grouped_report(&bytes, name, vb0) {
+            println!("{line}");
+        }
+    }
+}

crates/sylpheed-formats/src/mesh.rs

@@ -755,6 +755,60 @@ pub fn debug_try_anchor(
     None
 }
 
+/// Diagnostic: why does the decoder refuse a grouped-pool resource at a given
+/// pool start? Recomputes the pool layout exactly as [`anchor_grouped_meshes`]
+/// does and reports the pivot sub-mesh's verdict for each index/vertex pad —
+/// so a capture-proven pool that the decoder rejects names the gate to fix.
+pub fn debug_grouped_report(bytes: &[u8], name: &str, vb0: usize) -> Vec<String> {
+    let Some((decl, markers)) = decl_of(bytes, name) else {
+        return vec!["no such XBG7 resource".into()];
+    };
+    let n = markers.len();
+    if n == 0 {
+        return vec!["no index markers".into()];
+    }
+    let (mut rel_ib, mut acc_i) = (Vec::with_capacity(n), 0usize);
+    for &(_, ic) in &markers {
+        rel_ib.push(acc_i);
+        acc_i = align4(acc_i + ic * 2);
+    }
+    let span = rel_ib[n - 1] + markers[n - 1].1 * 2;
+    let kmax = (0..n).max_by_key(|&i| markers[i].1).unwrap_or(0);
+    let (vck, ick) = markers[kmax];
+    let mut off_v = 0usize;
+    for &(vc, _) in markers.iter().take(kmax) {
+        off_v += vc * decl.stride;
+    }
+    let mut out = vec![format!(
+        "{name}: {n} sub-meshes, pivot #{kmax} ({vck} verts, {ick} idx), pool span {span}"
+    )];
+    for pad in 0..=3usize {
+        if vb0 < span + pad {
+            out.push(format!("  pad {pad}: pool start is before the index pool"));
+            continue;
+        }
+        let ib0 = vb0 - span - pad;
+        // Same gates the production pivot test uses: strict winding (0.85)
+        // AND connectivity. Reporting at 0.0 would accept blocks the decoder
+        // rejects and send the reader chasing the wrong gate.
+        let verdict = validate_block_report(
+            bytes,
+            ib0 + rel_ib[kmax],
+            vb0 + off_v,
+            vck,
+            ick,
+            &decl,
+            0.85,
+            true,
+        );
+        out.push(match verdict {
+            Ok(()) => format!("  pad {pad}: ACCEPTED"),
+            Err(why) => format!("  pad {pad}: {why}"),
+        });
+    }
+    out
+}
+
 /// Diagnostic: for a resource whose vertex buffer is *known* (a runtime capture
 /// names it), where could its index buffer be? The anchor scan assumes the index
 /// buffer sits immediately before the vertex buffer; this scans the whole
@@ -998,17 +1052,43 @@ fn validate_block(
     min_consistency: f32,
     strict_connectivity: bool,
 ) -> bool {
+    validate_block_report(
+        bytes,
+        ib,
+        vb,
+        vtx_count,
+        index_count,
+        decl,
+        min_consistency,
+        strict_connectivity,
+    )
+    .is_ok()
+}
+
+/// [`validate_block`], but naming the gate that rejected a block. A runtime
+/// capture can prove a block is real; when the decoder still refuses it, this
+/// says which test is wrong rather than leaving a threshold to be guessed at.
+fn validate_block_report(
+    bytes: &[u8],
+    ib: usize,
+    vb: usize,
+    vtx_count: usize,
+    index_count: usize,
+    decl: &VertexDecl,
+    min_consistency: f32,
+    strict_connectivity: bool,
+) -> Result<(), String> {
     let stride = decl.stride;
     let idx_bytes = index_count * 2;
     if ib + idx_bytes > bytes.len() {
-        return false;
+        return Err("index buffer runs past the container".into());
     }
     let vtx_bytes = match vtx_count.checked_mul(stride) {
         Some(v) => v,
-        None => return false,
+        None => return Err("vertex byte count overflows".into()),
     };
     if vb + vtx_bytes > bytes.len() {
-        return false;
+        return Err("vertex buffer runs past the container".into());
     }
 
     // ── Full index validation: every index in range, uses ~all vertices. ──
@@ -1016,12 +1096,14 @@ fn validate_block(
     for k in 0..index_count {
         let i = be16(bytes, ib + k * 2) as u32;
         if i >= vtx_count as u32 {
-            return false;
+            return Err(format!("index {i} out of range (vtx_count {vtx_count})"));
         }
         max_idx = max_idx.max(i);
     }
     if (max_idx as usize) + 4 < vtx_count {
-        return false;
+        return Err(format!(
+            "indices reach only {max_idx} of {vtx_count} vertices (buffer not covered)"
+        ));
     }
 
     // ── Triangle quality: finite, non-degenerate, real spatial extent. ──
@@ -1046,7 +1128,7 @@ fn validate_block(
             for (a, slot) in pc.iter_mut().enumerate() {
                 let x = bef(bytes, base + a * 4);
                 if !x.is_finite() || x.abs() > 1.0e6 {
-                    return false;
+                    return Err(format!("position component {x} is not finite/plausible"));
                 }
                 *slot = x;
                 lo[a] = lo[a].min(x);
@@ -1088,7 +1170,9 @@ fn validate_block(
     }
     let extent = (hi[0] - lo[0]).max(hi[1] - lo[1]).max(hi[2] - lo[2]);
     if extent < 0.5 || sampled == 0 || degenerate * 10 > sampled * 3 {
-        return false; // too flat, or >30% degenerate → not this block
+        return Err(format!(
+            "extent {extent:.3} (min 0.5), {degenerate}/{sampled} degenerate (max 30%)"
+        ));
     }
     // Connectivity check: a correctly-anchored mesh has triangle edges that
     // are SMALL relative to its overall size (~0.05–0.15 of the bbox
@@ -1111,7 +1195,10 @@ fn validate_block(
         // `XBG7_EDGE_CAP_SMALL` instead — a targeted relaxation, off by default.
         let cap = if tris < small_tris() { small_cap() } else { edge_cap() };
         if mean_edge / diag > cap {
-            return false;
+            return Err(format!(
+                "connectivity: mean_edge/diag {:.3} > cap {cap:.2}",
+                mean_edge / diag
+            ));
         }
     }
     // Winding-consistency gate (same as `from_xpr2`): a correctly-anchored
@@ -1123,10 +1210,13 @@ fn validate_block(
     if nrm_counted > 0 && min_consistency > 0.0 {
         let na = nrm_agree as f32 / nrm_counted as f32;
         if na.max(1.0 - na) < min_consistency {
-            return false;
+            return Err(format!(
+                "winding consistency {:.3} < {min_consistency:.2}",
+                na.max(1.0 - na)
+            ));
         }
     }
-    true
+    Ok(())
 }
 
 /// Decode a **grouped-pool** XBG7 resource: one whose descriptor holds *several*

crates/sylpheed-formats/tests/mesh_consistency_disc.rs

@@ -108,10 +108,9 @@ fn shared_resources_decode_identically_in_every_container() {
 /// A runtime capture showed the container stores both halves of the `e106` hull
 /// as separate X-reflected buffers, so a `…_01`/`…_02` pair of equal vertex
 /// count should come out mirrored (or related by another axis / vertex order) —
-/// never identical, which is the collapse distinct assignment fixes. Disc-wide
+/// never identical, which is the collapse distinct assignment fixes. Since that
-/// this holds for every such pair except `n206`, whose twins are grouped-pool
+/// fix reached grouped pools too, this holds for **every** twin pair on the disc
-/// resources (4 sub-meshes) and so are excluded from distinct assignment; that
+/// — including `n206`, which was the last exception.
-/// one is the remaining known case, asserted explicitly so it cannot grow.
 #[test]
 fn twin_pairs_do_not_share_a_buffer() {
     let Some(root) = disc_root() else {
@@ -141,6 +140,5 @@ fn twin_pairs_do_not_share_a_buffer() {
             }
         }
     }
-    collapsed.retain(|c| !c.starts_with("n206_01"));
     assert!(collapsed.is_empty(), "twin pairs sharing one buffer: {collapsed:?}");
 }

docs/re/structures/xbg7-mesh.md

@@ -945,13 +945,25 @@ cannot regress).
 The runtime oracle is unchanged and cross-container inconsistency falls another
 10 %. The suite, including the disc- and ISO-gated tests, stays green.
 
-**It does *not* clear the `n206` collapse**, and that is informative: both twins
+**It clears the `n206` collapse too** — a correction to what was written here
-are grouped, the loser is re-placed past the taken pool, and no alternative pool
+first. `n206_02` now anchors at `0x342d984` instead of sharing `0x33b6e54` with
-**validates** — so it keeps the collided decode. `n206_02` is therefore the same
+its twin. The earlier "no alternative pool validates" reading was wrong twice
-class as `e106_eng_02_l` was before the cap moved: the correct block is rejected
+over: `mesh::debug_grouped_report` (`examples/why_rejected.rs`) shows that pool
-by the validator, not lost to selection. Its correct pool is one of
+**ACCEPTED at pad 0** under the production gates, and `0x342d984` is in the
-`0x342d984` (direct) or `0x33b7754` / `0x342e284` (mirrored); a capture of a
+candidate start list. What actually misled the check was the audit itself —
-stage containing `n206` would say which, and is the cheapest way to settle it.
+it classified twins by decoded **geometry**, and `0x342d984` is a *direct*
+(unmirrored) copy of `0x33b6e54`, so the pair still looked "identical" after it
+had been separated.
+
+The audit now distinguishes the two: identical geometry is only a collapse when
+it comes from **one buffer**. Re-run disc-wide, of 34 equal-count twin pairs:
+**18 exact X-mirror, 16 related another way, 0 identical-sharing-a-buffer, 0
+unrelated** — and the regression test no longer needs its `n206` exception.
+
+❔ Left open: whether `n206_01`/`n206_02` *should* be a mirrored pair at all. The
+container holds two direct copies **and** two mirrored ones
+(`0x33b7754`, `0x342e284`); our twins take the two direct copies, which is
+self-consistent but unverified — `n206` appears in no captured stage.
 
 ### The twin invariant, checked disc-wide (2026-08-12)