@@ -1,4 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+// test-budget: 20 — whole-package import: the in-process case imports the whole @objectstack/lint barrel cold and loads each lazy dep on its trigger, and the two built-dist probes each spawn node once to cold-load the dist and every lazy dep (typescript among them); slowest CI reading 13.91 s (run 40200279306, refresh PR #25960's window (hourly runs 40200279306 to 40272142430)) x 1.25, rounded up
 //
 // Boot-path contract: importing @objectstack/lint must NOT load its heavy,
 // gate-only dependencies. The package sits on the kernel boot path, while each
 
@@ -87,10 +87,50 @@ const NON_RECORD_ENTRIES: readonly (readonly [string, unknown])[] = [
 ];
 
 const stackWith = (junk: unknown): AnyRec => ({ objects: [junk, VALID_OBJECT] });
-const validOnly: AnyRec = { objects: [VALID_OBJECT] };
+const validOnly = (): AnyRec => ({ objects: [VALID_OBJECT] });
 
-const findingCount = (stack: AnyRec): number =>
-  AUTHORING_RULES.reduce((n, rule) => n + rule.run(stack, {}).length, 0);
+/** What one rule did with one stack: the message it threw, or how many findings it returned. */
+type RuleOutcome = { readonly rule: string } & ({ readonly threw: string } | { readonly findings: number });
+
+/**
+ * One sweep per distinct stack. Every rule is pure, `(stack) => Finding[]`, so
+ * running the table once over a stack and recording what each rule did answers
+ * both questions a case asks of that stack — did any rule throw, and how many
+ * findings did the rest return. The "throws" case and the "count" case of one
+ * stack read the same sweep, and a control stack shared by every junk shape is
+ * swept once. Each rule still gets a stack of its own, built fresh.
+ */
+const sweeps = new Map<string, readonly RuleOutcome[]>();
+const sweep = (key: string, build: () => AnyRec): readonly RuleOutcome[] => {
+  let outcomes = sweeps.get(key);
+  if (!outcomes) {
+    outcomes = AUTHORING_RULES.map((rule): RuleOutcome => {
+      try {
+        return { rule: rule.name, findings: rule.run(build(), {}).length };
+      } catch (e) {
+        return { rule: rule.name, threw: (e as Error).message };
+      }
+    });
+    sweeps.set(key, outcomes);
+  }
+  return outcomes;
+};
+
+/** `rule: message` for every rule that threw. */
+const throwsOf = (outcomes: readonly RuleOutcome[]): string[] =>
+  outcomes.flatMap((o) => ('threw' in o ? [`${o.rule}: ${o.threw}`] : []));
+
+/**
+ * The finding count over every rule not in `skip`. A rule outside `skip` that
+ * threw fails the count with its own message rather than counting as zero, as
+ * the count did when it ran the rules itself.
+ */
+const countOf = (outcomes: readonly RuleOutcome[], skip: readonly string[] = []): number =>
+  outcomes.reduce((n, o) => {
+    if (skip.includes(o.rule)) return n;
+    if ('threw' in o) throw new Error(`${o.rule}: ${o.threw}`);
+    return n + o.findings;
+  }, 0);
 
 describe('a non-record entry in `stack.objects` (#15552)', () => {
   /**
@@ -103,23 +143,17 @@ describe('a non-record entry in `stack.objects` (#15552)', () => {
     expect(AUTHORING_RULES.length).toBeGreaterThanOrEqual(42);
   });
 
-  describe.each(NON_RECORD_ENTRIES)('with %s beside a valid object', (_label, junk) => {
+  describe.each(NON_RECORD_ENTRIES)('with %s beside a valid object', (label, junk) => {
+    const withJunk = (): readonly RuleOutcome[] => sweep(`objects · ${label}`, () => stackWith(junk));
+
     it('no authoring rule throws', () => {
-      const threw: string[] = [];
-      for (const rule of AUTHORING_RULES) {
-        try {
-          rule.run(stackWith(junk), {});
-        } catch (e) {
-          threw.push(`${rule.name}: ${(e as Error).message}`);
-        }
-      }
-      expect(threw).toEqual([]);
+      expect(throwsOf(withJunk())).toEqual([]);
     });
 
     it('still judges the valid object standing beside it', () => {
-      const baseline = findingCount(validOnly);
+      const baseline = countOf(sweep('objects · valid only', validOnly));
       expect(baseline).toBeGreaterThan(0);
-      expect(findingCount(stackWith(junk))).toBe(baseline);
+      expect(countOf(withJunk())).toBe(baseline);
     });
   });
 
@@ -575,19 +609,12 @@ describe('a non-record entry in any other stack collection (#15636)', () => {
 
     describe.each(NON_RECORD_ENTRIES)('with %s', (shape, junk) => {
       const key = `${collection.label} · ${shape}`;
+      const withJunk = (): readonly RuleOutcome[] => sweep(key, () => collection.stack(members(junk)));
 
       it('throws out of no rule but the ones still filed as broken', () => {
-        const threw: string[] = [];
-        const detail: string[] = [];
-        for (const rule of AUTHORING_RULES) {
-          try {
-            rule.run(collection.stack(members(junk)), {});
-          } catch (e) {
-            threw.push(rule.name);
-            detail.push(`${rule.name}: ${(e as Error).message}`);
-          }
-        }
-        expect(threw.sort(), detail.join(' | ')).toEqual([...(RESIDUAL_THROWS[key] ?? [])].sort());
+        const outcomes = withJunk();
+        const threw = outcomes.flatMap((o) => ('threw' in o ? [o.rule] : []));
+        expect(threw.sort(), throwsOf(outcomes).join(' | ')).toEqual([...(RESIDUAL_THROWS[key] ?? [])].sort());
       });
 
       it('invents no finding about the entry no author wrote', () => {
@@ -596,9 +623,8 @@ describe('a non-record entry in any other stack collection (#15636)', () => {
         // folding that into the population count would let a crash read as
         // "reported nothing", which is the confusion this file exists to end.
         const skip = RESIDUAL_THROWS[key] ?? [];
-        const count = (stack: AnyRec): number =>
-          AUTHORING_RULES.reduce((n, rule) => (skip.includes(rule.name) ? n : n + rule.run(stack, {}).length), 0);
-        expect(count(collection.stack(members(junk))) - count(control())).toBe(RESIDUAL_INVENTED[key] ?? 0);
+        const controlOutcomes = sweep(`${collection.label} · control`, control);
+        expect(countOf(withJunk(), skip) - countOf(controlOutcomes, skip)).toBe(RESIDUAL_INVENTED[key] ?? 0);
       });
     });
   });
 
@@ -1,4 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+// test-budget: 20 — whole-package import: the in-process cases import the @objectstack/lint/runtime entry cold, which reaches the whole gate graph, and the built-dist probes each spawn node once to cold-load it; slowest CI reading 13.05 s (run 40200279306, refresh PR #25960's window (hourly runs 40200279306 to 40272142430)) x 1.25, rounded up
 //
 // The kernel boot-path contract for `@objectstack/lint/runtime` (#4463).
 //
 
@@ -1,5 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
-// test-budget: 140 — spawns the real producer once per single-cause fixture: every case runs the real schema generator (build-schemas.ts --check) in its own sandbox git tree, one fixture per cause; measured 2026-10-23 on its first CI reading after #25372 split build-schemas-check-mode.test.ts (run 39963105499, Test Core, head dce4c6406): 104.13 s x 1.25, rounded up; one reading only, its wall tracks the run's load, so re-measure once the window has more
+// test-budget: 190 — spawns the real producer once per single-cause fixture: every case runs the real schema generator (build-schemas.ts --check) in its own sandbox git tree, one fixture per cause, and one generator run is about 7 s of JSON Schema emission and spec module load before any assertion; re-measured 2026-10-26 over the 10 executed readings in refresh PR #25960's window (hourly runs 40200279306 to 40272142430), slowest 144.97 s (run 40204927516) x 1.25, rounded up
 //
 // The in-tree anchor: it carries the deletion gate offline (#5235), and only
 // --update-base moves it (#5358).
 
@@ -1,5 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
-// test-budget: 180 — spawns the real producer once per single-cause fixture: every case runs the real schema generator (build-schemas.ts --check) in its own sandbox git tree, one fixture per cause; measured 2026-10-23 on its first CI reading after #25372 split build-schemas-check-mode.test.ts (run 39963105499, Test Core, head dce4c6406): 140.73 s x 1.25, rounded up; one reading only, its wall tracks the run's load, so re-measure once the window has more
+// test-budget: 290 — spawns the real producer once per single-cause fixture: every case runs the real schema generator (build-schemas.ts --check) in its own sandbox git tree, one fixture per cause, and one generator run is about 7 s of JSON Schema emission and spec module load before any assertion; re-measured 2026-10-26 over the 10 executed readings in refresh PR #25960's window (hourly runs 40200279306 to 40272142430), slowest 226.64 s (run 40204927516) x 1.25, rounded up
 //
 // The manifest ratchet and its neighbours: a check reports, it does not write (#4711);
 // an export that never published must be declared (#16431); the output clean spares a
@@ -557,7 +557,7 @@ describe('build-schemas.ts — the output clean spares a sibling generator (#537
 describe('build-schemas.ts — the bundle counts the definitions it carries (#12588)', () => {
   const OUT = () => path.join(sandbox, 'json-schema');
 
-  beforeEach(() => {
+  beforeAll(() => {
     // A current, self-consistent tree, so the run exits 0 and these assertions
     // are about the bundle rather than about some ratchet upstream of it.
     seedManifest((s) => s);
@@ -566,11 +566,20 @@ describe('build-schemas.ts — the bundle counts the definitions it carries (#12
     seedSurfaceBase(tip, (k) => k);
   });
 
+  /**
+   * One generation for the block. Both cases read the same fixture and only
+   * READ what the run printed and wrote, so the first case to need it runs the
+   * generator and the other reads that run, the way
+   * `openapi-self-consistency.test.ts` shares its shipped run.
+   */
+  let generation: ReturnType<typeof run> | undefined;
+  const generated = (): ReturnType<typeof run> => (generation ??= run([]));
+
   it(
     'writes x-schema-count equal to its own $defs size, and to the files on disk',
     { timeout: SPAWN_TIMEOUT_MS },
     () => {
-      const { status, output } = run([]);
+      const { status, output } = generated();
       expect(status).toBe(0);
 
       const bundle = JSON.parse(
@@ -605,7 +614,7 @@ describe('build-schemas.ts — the bundle counts the definitions it carries (#12
     'accounts for every emit the definition count does not include',
     { timeout: SPAWN_TIMEOUT_MS },
     () => {
-      const { status, output } = run([]);
+      const { status, output } = generated();
       expect(status).toBe(0);
 
       const bundle = JSON.parse(
 
@@ -1,5 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
-// test-budget: 190 — spawns the real producer once per single-cause fixture: every case runs the real schema generator (build-schemas.ts --check) in its own sandbox git tree, one fixture per cause; measured 2026-10-23 on its first CI reading after #25372 split build-schemas-check-mode.test.ts (run 39963105499, Test Core, head dce4c6406): 147.5 s x 1.25, rounded up; one reading only, its wall tracks the run's load, so re-measure once the window has more
+// test-budget: 310 — spawns the real producer once per single-cause fixture: every case runs the real schema generator (build-schemas.ts --check) in its own sandbox git tree, one fixture per cause, and one generator run is about 7 s of JSON Schema emission and spec module load before any assertion; re-measured 2026-10-26 over the 10 executed readings in refresh PR #25960's window (hourly runs 40200279306 to 40272142430), slowest 241.05 s (run 40204927516) x 1.25, rounded up
 //
 // RETIRED_KEYS_BY_MAJOR rows, judged by exact key: check (b) matches the retired key,
 // not its leaf (#4659); a nested retirement row is judged, not ignored (#17969).
 
@@ -1,4 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+// test-budget: 20 — spawns the real producer once per single-cause fixture: the subprocess group runs the real build-openapi.ts once per distinct fixture, the shipped source and each mutation, and cases that share a fixture read its one run; slowest CI reading 10.48 s (run 40204927516, refresh PR #25960's window (hourly runs 40200279306 to 40272142430)) x 1.25, rounded up
 //
 // Pins the self-consistency gate on the generated OpenAPI document (#5168).
 //
 
@@ -1,4 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+// test-budget: 20 — spawns the real producer once per single-cause fixture: each build case runs the real tsup build (in-process, with the package's own main options and its esbuild plugin) once over the entry or fixture it pins, and the ./shared case bundles two consumers of that output; slowest CI reading 11.86 s (run 40214410982, refresh PR #25960's window (hourly runs 40200279306 to 40272142430)) x 1.25, rounded up
 //
 // The `pure-schema-construction` build plugin (`tsup.config.ts`) marks real
 // calls only (#20686). Its earlier line-prefix rule also rewrote a marked name
 
@@ -1,4 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
+// test-budget: 20 — compiles the package type graph once: the consumer program is one tsc run over the BUILT root entry's declarations, and the canary is one more tsc run over a one-line fixture; slowest CI reading 12.92 s (run 40214410982, refresh PR #25960's window (hourly runs 40200279306 to 40272142430)) x 1.25, rounded up
 
 /**
  * Root-entry nameability pin (#11350) — the consumer shape, verbatim.
 
@@ -1,5 +1,5 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
-// test-budget: 30 — spawns the real producer once per single-cause fixture: the last block spawns the real build-upgrade-guide.ts once per mode or usage error it pins (twice for the determinism pin), and the rest are in-process; slowest CI reading 19.7 s (hourly run 38798542648) x 1.25, rounded up
+// test-budget: 40 — spawns the real producer once per single-cause fixture: the last block spawns the real build-upgrade-guide.ts once per mode or usage error it pins (twice for the determinism pin), and the rest are in-process; re-measured 2026-10-26 over the 8 executed readings in refresh PR #25960's window (hourly runs 40200279306 to 40272142430), slowest 30.07 s (run 40214410982) x 1.25, rounded up
 //
 // Pins the protocol upgrade guide's public address (#22449 B′, condition 2):
 // one docs-site page per protocol major plus an index, generated at docs build
Original file line number	Diff line number	Diff line change
`@@ -1,4 +1,5 @@`
`1`	`1`	`// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.`
	`2`	`+// test-budget: 20 — whole-package import: the in-process case imports the whole @objectstack/lint barrel cold and loads each lazy dep on its trigger, and the two built-dist probes each spawn node once to cold-load the dist and every lazy dep (typescript among them); slowest CI reading 13.91 s (run 40200279306, refresh PR #25960's window (hourly runs 40200279306 to 40272142430)) x 1.25, rounded up`
`2`	`3`	`//`
`3`	`4`	`// Boot-path contract: importing @objectstack/lint must NOT load its heavy,`
`4`	`5`	`// gate-only dependencies. The package sits on the kernel boot path, while each`
Original file line number	Diff line number	Diff line change
`@@ -1,4 +1,5 @@`
`1`	`1`	`// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.`
	`2`	`+// test-budget: 20 — spawns the real producer once per single-cause fixture: the subprocess group runs the real build-openapi.ts once per distinct fixture, the shipped source and each mutation, and cases that share a fixture read its one run; slowest CI reading 10.48 s (run 40204927516, refresh PR #25960's window (hourly runs 40200279306 to 40272142430)) x 1.25, rounded up`
`2`	`3`	`//`
`3`	`4`	`// Pins the self-consistency gate on the generated OpenAPI document (#5168).`
`4`	`5`	`//`
Original file line number	Diff line number	Diff line change
`@@ -1,4 +1,5 @@`
`1`	`1`	`// Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.`
	`2`	`+// test-budget: 20 — compiles the package type graph once: the consumer program is one tsc run over the BUILT root entry's declarations, and the canary is one more tsc run over a one-line fixture; slowest CI reading 12.92 s (run 40214410982, refresh PR #25960's window (hourly runs 40200279306 to 40272142430)) x 1.25, rounded up`
`2`	`3`
`3`	`4`	`/**`
`4`	`5`	`* Root-entry nameability pin (#11350) — the consumer shape, verbatim.`