feat(ci): gate release workflow on resolution precision/recall thresholds (#886)

* feat(ci): gate release workflow on resolution precision/recall thresholds (#875)

Add resolution quality gates to the benchmark pipeline so regressions
are caught before publishing:

- benchmark.yml: run vitest resolution test after the benchmark script,
  failing the workflow if any language drops below its threshold
- update-benchmark-report.ts: warn on precision >5pp or recall >10pp
  drop per language between releases
- regression-guard.test.ts: hard-fail CI on precision/recall regressions
  across releases, with KNOWN_REGRESSIONS exemption support

* style: fix biome formatting in regression guard

* fix: add SYNC comments for duplicated thresholds and eliminate redundant file read (#886)

Add cross-reference SYNC comments between regression-guard.test.ts and
update-benchmark-report.ts so the duplicated precision/recall thresholds
stay in lockstep. Replace the second extractJsonData call with a type
cast of buildHistory since both read the same file and marker.

* fix: add timeout-minutes to resolution gate step (#886)

Prevents a hanging WASM build from stalling the entire benchmark job
indefinitely. 30-minute cap is generous enough for the full language
fixture suite while still bounding worst-case CI time.

* fix: document resolution key format in KNOWN_REGRESSIONS comment (#886)

Author: carlos-alm

.github/workflows/benchmark.yml

@@ -108,6 +108,7 @@ jobs:
 
       - name: Gate on resolution thresholds
         if: steps.existing.outputs.skip != 'true'
+        timeout-minutes: 30
         run: npx vitest run tests/benchmarks/resolution/resolution-benchmark.test.ts --reporter=verbose
 
       - name: Merge resolution into build result

scripts/update-benchmark-report.ts

@@ -322,6 +322,8 @@ if (prev) {
 
 // ── Resolution regression detection ─────────────────────────────────────
 // Resolution metrics are "higher is better" — warn when they DROP.
+// SYNC: These must match PRECISION_DROP_PP / RECALL_DROP_PP in
+// tests/benchmarks/regression-guard.test.ts (the hard-fail gate side).
 const PRECISION_DROP_THRESHOLD = 0.05; // warn if precision drops >5pp
 const RECALL_DROP_THRESHOLD = 0.10;    // warn if recall drops >10pp
 

tests/benchmarks/regression-guard.test.ts

@@ -61,6 +61,7 @@ const SKIP_VERSIONS = new Set(['3.8.0']);
  * underlying issue is being fixed.
  *
  * Format: "version:metric-label" (must match the label passed to checkRegression).
+ * Resolution keys use: "version:resolution <lang> precision" or "version:resolution <lang> recall".
  *
  * - 3.9.0:1-file rebuild — native incremental path re-runs graph-wide phases
  *   (structureMs, AST, CFG, dataflow) on single-file rebuilds. Documented in
@@ -521,6 +522,9 @@ describe('Benchmark regression guard', () => {
      * Precision >5pp drop and recall >10pp drop are flagged.
      * Recall has a wider threshold because it's more volatile — adding new
      * expected edges to fixtures can temporarily lower recall.
+     *
+     * SYNC: These must match PRECISION_DROP_THRESHOLD / RECALL_DROP_THRESHOLD
+     * in scripts/update-benchmark-report.ts (the ::warning annotation side).
      */
     const PRECISION_DROP_PP = 0.05;
     const RECALL_DROP_PP = 0.1;
@@ -539,10 +543,9 @@ describe('Benchmark regression guard', () => {
       resolution?: Record<string, ResolutionLang>;
     }
 
-    const fullHistory = extractJsonData<BuildEntryWithResolution>(
-      path.join(BENCHMARKS_DIR, 'BUILD-BENCHMARKS.md'),
-      'BENCHMARK_DATA',
-    );
+    // buildHistory already parsed BUILD-BENCHMARKS.md with the same marker;
+    // widen the type instead of re-reading the file.
+    const fullHistory = buildHistory as BuildEntryWithResolution[];
 
     const resolutionPair = findLatestPair(fullHistory, (e) => e.resolution != null);