feat(native): make native model analysis the default with a native analysis cache

.github/workflows/verify.yml

@@ -322,7 +322,7 @@ jobs:
         run: >-
           make compiler-differential-shipped
           COMPILER_DIFFERENTIAL_ARGS="--stage-captures --require-discovery-coverage
-          --require-render-coverage
+          --require-render-coverage --require-analysis-coverage
           --evidence-dir ${{ runner.temp }}/compiler-differential-evidence/shipped"
       - name: Upload differing projects' targets, stderr and compiled-test stats
         if: failure()
@@ -606,25 +606,27 @@ jobs:
         include:
           - kind: dense
             max_ratio: "1.10"
+            edit_max_ratio: "1.20"
+            edit_max_cpu_ratio: "1.75"
           - kind: fresh
             max_ratio: "1.10"
           - kind: dense
-            label: native-preview contracts vs python
+            label: native contracts vs python
             base_engine: python
-            head_engine: native-preview
+            head_engine: native
             gate_phase: contracts_cpu_ms
             max_ratio: "1.00"
           - kind: dense
-            label: native-preview model analysis vs python
+            label: native model analysis vs python
             base_engine: python
-            head_engine: native-preview
+            head_engine: native
             gate_phase: model_analysis_cpu_ms
             compile_args: --compile-arg=--select=*
             max_ratio: "1.50"
           - kind: dense
-            label: native-preview custom rules vs python
+            label: native custom rules vs python
             base_engine: python
-            head_engine: native-preview
+            head_engine: native
             gate_phase: custom_rules_cpu_ms
             max_ratio: "1.25"
     steps:
@@ -653,7 +655,7 @@ jobs:
       - name: Install base dependencies
         working-directory: base
         run: uv sync --locked --all-groups --all-extras
-      - name: Compare python and native-preview phase CPU on head
+      - name: Compare python and native phase CPU on head
         if: matrix.gate_phase
         env:
           KIND: ${{ matrix.kind }}
@@ -679,10 +681,14 @@ jobs:
         env:
           KIND: ${{ matrix.kind }}
           MAX_RATIO: ${{ matrix.max_ratio }}
+          EDIT_MAX_RATIO: ${{ matrix.edit_max_ratio }}
+          EDIT_MAX_CPU_RATIO: ${{ matrix.edit_max_cpu_ratio }}
         run: |
           uv run python -m scripts.compare_compile_performance \
             --kind "${KIND}" \
             --modes cold warm edit \
+            ${EDIT_MAX_RATIO:+--edit-max-ratio "$EDIT_MAX_RATIO"} \
+            ${EDIT_MAX_CPU_RATIO:+--edit-max-cpu-ratio "$EDIT_MAX_CPU_RATIO"} \
             --max-ratio "${MAX_RATIO}" \
             --models 3000 \
             --base-root base \

AGENTS.md

@@ -55,6 +55,16 @@
 - Performance fixes must not reduce coverage. Show that diagnostics are unchanged or a strict,
   validated superset on the dense benchmark and a real-project run.
 - Do not raise performance budgets or tolerances to make a regression pass.
+- Temporary user-approved exception (2026-10-10), removable after post-conversion memory
+  optimisation: only the 10,000-model case in `test_fresh_process_compile_cache_performance.py`
+  may use 3 GiB peak RSS instead of 2.5 GiB; release-performance `lineage column trace` may use
+  a 2.0 peak RSS ratio and `dag --json` a 1.5 ratio instead of 1.25. Smaller cache cases,
+  cold-only guards, all other commands, and all speed/correctness limits retain their budgets.
+  The prohibition above applies to every other budget or tolerance increase.
+- Additional temporary user-approved exception (2026-10-10): the dense 3,000-model
+  base/head edit comparison may use 1.20x wall and 1.75x CPU; release-performance
+  `compile (one-model edit)` may use 1.35x peak RSS. Remove after post-conversion
+  optimisation. Cold/warm modes, other workloads and correctness gates remain unchanged.
 
 ## Native SQL (Polyglot) Changes
 

Makefile

@@ -304,7 +304,7 @@ compiler-differential:
 		$(COMPILER_DIFFERENTIAL_GATES) $(COMPILER_DIFFERENTIAL_ARGS)
 
 # python vs native-preview above covers every native stage; this keeps the shipped `native` default
-# (native discovery and rendering) and its goldens covered on the same corpus.
+# (native discovery, rendering and analysis) and its goldens covered on the same corpus.
 compiler-differential-shipped:
 	env PYTHONUNBUFFERED=1 uv run python -m scripts.run_compiler_differential \
 		--engines python native --corpus fixtures examples seeds failures \
@@ -319,7 +319,7 @@ compiler-baselines:
 	$(MAKE) compiler-differential COMPILER_DIFFERENTIAL_ARGS="$(COMPILER_DIFFERENTIAL_CI_ARGS) \
 		--require-analysis-coverage --native-fallbacks update --goldens update"
 	$(MAKE) compiler-differential-shipped COMPILER_DIFFERENTIAL_ARGS="$(COMPILER_DIFFERENTIAL_CI_ARGS) \
-		--native-fallbacks update --goldens update"
+		--require-analysis-coverage --native-fallbacks update --goldens update"
 
 compiler-differential-dense:
 	env PYTHONUNBUFFERED=1 uv run python -m scripts.run_compiler_differential \

crates/README.md

@@ -62,9 +62,10 @@ stages run natively:
 | `native` | The default: native stages that passed their flip gate (`shipped` tier). |
 | `native-preview` | Opt-in: shipped stages plus stages still in development (`preview` tier). |
 
-The shipped tier covers discovery and rendering: declaration files and scopes, model config,
-reference extraction, the model loop, macro calls and the macro-call store, and attachments. SQL
-analysis, contracts, lineage, SQL-test glue and project assembly are still `preview`. A failing
+The shipped tier covers discovery, rendering and analysis: declaration files and scopes, model
+config, reference extraction, the model loop, macro calls and the macro-call store, attachments,
+the type system, model analysis and its `model-analyses.bin` cache, semantic checks, contracts,
+lineage facts, SQL-test glue and project assembly. A failing
 render raises its first error directly, as the Python stage would, and runs each macro call at
 most once; a divergence the bridge detects is reported as a native stage mismatch. Because the
 macro-call store is shipped, the

crates/sqlbuild-analysis/Cargo.toml

@@ -20,6 +20,7 @@ regex.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 sha2.workspace = true
+sqlbuild-cache.workspace = true
 sqlbuild-core.workspace = true
 sqlbuild-model-config.workspace = true
 sqlbuild-scopes.workspace = true