fix: keep fully cold large compiles within the memory budget

Author: kvlonge

docs/dense-compile-performance.md

@@ -26,7 +26,7 @@ The guard checks:
 | 1,000 | 14s | 1.5GiB |
 | 3,000 | 34s | 2.75GiB |
 | 5,000 | 55s | 3.25GiB |
-| 10,000 | 115s | 4GiB |
+| 10,000 | 145s | 4GiB |
 
 These are regression ceilings for the implemented compiler, with headroom for
 host variation. They are not a claim of sub-10-second compilation at every size.
@@ -47,8 +47,10 @@ analysis results in batches of 64. Smaller requests retain a single batch. Both
 use up to four workers; local resolution views avoid copying CTE definitions
 that are already available through the source catalogue.
 These limits change scheduling, not analysis or fallback semantics.
-Projects with at least 10,000 model projections use two native analysis workers to
+Projects with at least 10,000 model projections use one native analysis worker to
 leave memory headroom for the isolated custom-rule host after analysis.
+The 10k cold ceiling includes headroom for this memory-bounded scheduling; warm
+cache hits and small edit batches keep their existing scheduling.
 
 The compiler reuses borrowed syntax-tree facts when it can prove complete reference
 binding and infer types without consulting mutable type annotations. Unsupported

src/sqlbuild/compiler/compile/_helpers/analysis/compact.py

@@ -626,7 +626,7 @@ def _prepare_compact_analysis_batch(
 def _run_compact_analysis_batch(*, preparation: CompactBatchPreparation) -> object:
     """Bound resident analysis heaps before the isolated custom-rule host starts."""
 
-    workers: int = 2 if len(preparation.projections) >= _LARGE_COMPACT_PROJECT_MODELS else 4
+    workers: int = 1 if len(preparation.projections) >= _LARGE_COMPACT_PROJECT_MODELS else 4
     return orjson.loads(
         cast(NativeQueryAnalysisModule, _native).analyze_project_queries_compact_json(
             orjson.dumps(

tests/e2e/src/sqlbuild/cli/commands/main/compile/test_dense_compile_performance.py

@@ -59,7 +59,7 @@
         DenseCompileGuardTestCase(
             "dense_models_10000_all_rules",
             10000,
-            115.0,
+            145.0,
             4 * _GIB,
             "8125ddde6d3ae172bb15878f95048d66398453c516bbd837b1fd83c04c5982b5",
         ),