fix(scene): unfreeze open on huge xref Inserts

A real DWG with an xref containing ~74k sub-entities froze on the first
render: Insert tessellation exploded the block, style-resolved every sub
serially, and one bad sub could hang the loop forever.

- Insert short-circuits to an insertion marker when the referenced block
  has more than INSERT_SUB_LIMIT (5_000) sub-entities. Documented as
  temporary in RENDERING_OPTIMIZATION.md.
- Re-run purge_corrupt_entities after resolve_xrefs so parser-garbage
  inside xref content is caught (initial purge runs before xref load).
- build_derived_caches: single pass over entities collecting handles
  per kind; tessellation reads via doc.get_entity instead of cloning
  every EntityType. Image decode now parallel too.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

RENDERING_OPTIMIZATION.md

@@ -272,4 +272,42 @@ Phase 4.2  Hi-Z occlusion                   high complexity, 3D only, last
   proportional to visible fraction.
 - **Phase 2 target:** Pan/zoom frame cost O(visible) not O(total).
 - **Phase 3 target:** Complex solid scene at far zoom → mesh triangle count < 10% of full detail.
+
+---
+
+## TEMPORARY: Insert sub-entity count limit (revert when better fix lands)
+
+`tessellate_entity` in `src/scene/mod.rs` has a guard that short-circuits any
+Insert whose referenced block contains more than `INSERT_SUB_LIMIT` (currently
+5,000) sub-entities. The Insert is rendered as an insertion marker only — no
+explode, no sub tessellation.
+
+**Why this exists:** A real-world DWG was found where an xref block contained
+~74k sub-entities. Opening the file froze the UI on the first render because
+`tessellate_entity` exploded the block and serially style-resolved every
+sub-entity (and any single bad sub could hang tessellation forever). The guard
+made the file openable.
+
+**Why it should go away:** The limit hides legitimate geometry. Large xrefs are
+common in civil/infrastructure drawings — the user shouldn't have to see them
+as crosses.
+
+**Proper fixes that would let us drop this guard:**
+1. Per-block tessellation cache — tessellate a block once, then per-Insert
+   apply the transform to cached wire vertices. Helps when the same block is
+   inserted multiple times; doesn't fully solve single-Insert huge blocks.
+2. Lazy / background Insert tessellation — show marker first, fill in detail
+   on a background thread, redraw when ready.
+3. Per-sub-entity hardening — find why a single sub can hang tessellation
+   (likely Spline NURBS evaluation or malformed boundary edge) and fix the
+   root cause; combine with sub-count budget per frame instead of a hard cap.
+4. xref attach metadata respected — DWG carries an "unloaded" state for xref
+   blocks; honour it instead of always inlining xref content on open.
+
+When any of (1)–(3) lands, delete the `INSERT_SUB_LIMIT` block in
+`tessellate_entity` (search for `INSERT_SUB_LIMIT`).
+
+Related companion fix in `src/app/update.rs`: after `resolve_xrefs`, a second
+`purge_corrupt_entities` pass runs against the newly inlined xref content.
+That one is not a hack — it should stay even after the Insert guard goes.
 - **Phase 4 target:** GPU-cull overhead < 0.5 ms for 1M entity scene.