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>

docs/re/structures/xbg7-mesh.md

@@ -1113,11 +1113,24 @@ always runs over the whole container, and a subset is now a subset of the
 container's own answer. Verified: one-name, three-name and full decodes now
 return the identical offset for the same resource.
 
-**Cost, stated plainly:** a single-resource query on a 50 MB container went from
-near-instant to **~15 s**, because it now anchors every resource. The viewer
-calls this once per ship view (with cancellation), so it is a slow prepare step
-rather than a per-frame cost — but caching the per-container decode is the
-obvious follow-up and is not done here.
+**Cost, and the cache that pays it back.** Because the assignment must see every
+resource, a single-resource query became as expensive as a full decode (~15 s on
+a 50 MB container). `full_decode_cached` memoises the whole-container decode,
+keyed by a fingerprint of the bytes (length + three sampled 4 KB windows) and the
+consistency setting, holding the last four containers. Decoding five ships from
+`Stage_S02` in turn, as the viewer does:
+
+```
+e007: 2 models in 10.494 s     ← the one full decode
+e010: 2 models in    48.6 µs
+e105: 37 models in  1.42 ms
+e106: 32 models in  92.7 µs
+e108: 13 models in  73.7 µs
+```
+
+So a stage costs one decode, not one per ship. The whole-container path
+(`anchor_models_cancellable`, what the sweeps use) stays uncached — it is already
+the thing being measured.
 
 ### The consistency figure is mostly bounding boxes — real disagreement is ONE resource