perf(mesh): memoise the whole-container decode

Filtering after the assignment made every subset query a full decode (~15s on a
50MB container). full_decode_cached memoises it per container -- fingerprint is
length plus three sampled 4KB windows, keyed with min_consistency, last four kept.
Decoding five ships from Stage_S02 in turn: 10.5s for the first, then 48us-1.4ms.
A stage now costs one decode rather than one per ship. Ten suites green, viewer
builds, and a spot-checked resource still lands on the same offset.

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

crates/sylpheed-formats/examples/bench_ships.rs

@@ -0,0 +1,22 @@
+//! Time repeated per-ship decodes of one container, as the viewer does.
+use std::collections::HashSet;
+use std::time::Instant;
+use sylpheed_formats::mesh::Xbg7Model;
+use sylpheed_formats::ship::{is_base_part, ship_id_of};
+fn main() {
+    let a: Vec<String> = std::env::args().collect();
+    let bytes = std::fs::read(&a[1]).unwrap();
+    let names = sylpheed_formats::mesh::xbg7_resource_names(&bytes);
+    let ids: Vec<String> = {
+        let mut v: Vec<String> = names.iter().filter(|n| is_base_part(n))
+            .filter_map(|n| ship_id_of(n).map(|s| s.to_string())).collect();
+        v.sort(); v.dedup(); v.truncate(5); v
+    };
+    for id in &ids {
+        let want: HashSet<String> = names.iter()
+            .filter(|n| ship_id_of(n) == Some(id.as_str())).cloned().collect();
+        let t = Instant::now();
+        let got = Xbg7Model::models_named(&bytes, &want, &|| false);
+        println!("{id}: {} models in {:?}", got.len(), t.elapsed());
+    }
+}

crates/sylpheed-formats/src/mesh.rs

@@ -463,11 +463,32 @@ impl Xbg7Model {
         Self::anchor_models_filtered(bytes, min_consistency, should_cancel, None)
     }
 
+    /// Decode the whole container (cached), then hand back the requested subset.
+    ///
+    /// The decode itself must always see every resource — distinct assignment
+    /// resolves collisions against the whole population, and pruning first made
+    /// the answer depend on the request (see the module history). That makes a
+    /// single-resource query as expensive as a full decode, so the full decode is
+    /// memoised per container: the viewer asks for one ship's parts at a time and
+    /// would otherwise re-anchor 6 000 resources per ship.
     fn anchor_models_filtered(
         bytes: &[u8],
         min_consistency: f32,
         should_cancel: &(dyn Fn() -> bool + Sync),
         wanted: Option<&std::collections::HashSet<String>>,
+    ) -> Vec<Xbg7Model> {
+        let Some(w) = wanted else {
+            return Self::anchor_models_uncached(bytes, min_consistency, should_cancel, None);
+        };
+        let full = full_decode_cached(bytes, min_consistency, should_cancel);
+        full.iter().filter(|m| w.contains(&m.name)).cloned().collect()
+    }
+
+    fn anchor_models_uncached(
+        bytes: &[u8],
+        min_consistency: f32,
+        should_cancel: &(dyn Fn() -> bool + Sync),
+        wanted: Option<&std::collections::HashSet<String>>,
     ) -> Vec<Xbg7Model> {
         let mut out = Vec::new();
         if bytes.len() < 16 || &bytes[..4] != b"XPR2" {
@@ -1138,6 +1159,49 @@ pub fn debug_resource_params(bytes: &[u8], name: &str) -> Option<(Vec<(usize, us
     None
 }
 
+/// Memoised whole-container decode, keyed by a cheap fingerprint of the bytes
+/// plus the consistency setting. Holds the last few containers; a stage decode is
+/// a handful of MB, and the alternative is re-anchoring every resource for every
+/// ship the viewer shows.
+fn full_decode_cached(
+    bytes: &[u8],
+    min_consistency: f32,
+    should_cancel: &(dyn Fn() -> bool + Sync),
+) -> std::sync::Arc<Vec<Xbg7Model>> {
+    use std::sync::{Arc, Mutex, OnceLock};
+    // Fingerprint: length plus three sampled 4 KB windows. Two different
+    // containers agreeing on all of that is not a case this format produces.
+    let mut fp: u64 = 0xcbf2_9ce4_8422_2325 ^ bytes.len() as u64;
+    let windows = [0usize, bytes.len() / 2, bytes.len().saturating_sub(4096)];
+    for w in windows {
+        for b in bytes.iter().skip(w).take(4096) {
+            fp = (fp ^ *b as u64).wrapping_mul(0x100_0000_01b3);
+        }
+    }
+    let key = (fp, min_consistency.to_bits());
+    #[allow(clippy::type_complexity)]
+    static CACHE: OnceLock<Mutex<Vec<((u64, u32), Arc<Vec<Xbg7Model>>)>>> = OnceLock::new();
+    let cache = CACHE.get_or_init(|| Mutex::new(Vec::new()));
+    if let Ok(c) = cache.lock() {
+        if let Some((_, v)) = c.iter().find(|(k, _)| *k == key) {
+            return Arc::clone(v);
+        }
+    }
+    let models = Arc::new(Xbg7Model::anchor_models_uncached(
+        bytes,
+        min_consistency,
+        should_cancel,
+        None,
+    ));
+    if let Ok(mut c) = cache.lock() {
+        c.push((key, Arc::clone(&models)));
+        if c.len() > 4 {
+            c.remove(0);
+        }
+    }
+    models
+}
+
 /// Diagnostic: the candidate vertex-buffer starts the stage anchor scan will
 /// consider for a given `stride`, for one container. A runtime capture names the
 /// offsets the engine really drew from (see `examples/shared_vbase_check.rs`), so