fix(packs): pack test counts truth sites listed as folded evidence as covered

.changeset/score-folded-evidence.md

@@ -0,0 +1,5 @@
+---
+"uptide": patch
+---
+
+`uptide pack test` counts a ground-truth site that a folded finding lists as its evidence as covered by that finding, so recall reflects what `check` reports. The typescript pack's breaking recall goes from 17% to 99% on unchanged ground truth: its TypeScript 6 `types` default errors fold into one tsconfig finding each.

README.md

@@ -70,7 +70,7 @@ The contract: [migration packs](docs/packs.md).
 | `@sentry/nextjs` | 10.x → 11.x | 100% | 100% | 8 | cohstats/coh3-stats, xdoubleu/tools.xdoubleu.com, maximunited/yomu | verified |
 | `@sentry/react` | 10.x → 11.x | 100% | 100% | 2 | owine/youth-activity-scheduler, successible/cleanslate | candidate |
 | `stripe` | 14 and newer | 100% | 92% | 21 | unkeyed/unkey, nextjs/saas-starter, starl3xx/wallet-to-social, AtlasDevHQ/atlas | verified |
-| `typescript` | 5.x → 6.x | 100% | 17% | 1399 | m-lyon/recipe, robrechtme/recipes, JosephMaynard/dependency-radar, CroffZ/github-release-notifier, SignalK/signalk-derived-data, shogo82148/typescript-action, Borewit/music-metadata | verified |
+| `typescript` | 5.x → 6.x | 100% | 99% | 1399 | m-lyon/recipe, robrechtme/recipes, JosephMaynard/dependency-radar, CroffZ/github-release-notifier, SignalK/signalk-derived-data, shogo82148/typescript-action, Borewit/music-metadata | verified |
 | `vitest` | 4.x → 5.x | 100% | 50% | 32 | votingworks/vxsuite, vitorvasc/opentelemetry-ecosystem-explorer | verified |
 | `zod` | 3.x → 4.x | 100% | 64% | 219 | ugurkocde/AwesomeIntune, howardyang2009/PATH, band-ai/band-sdk-typescript, vzdc-artcc/website, CANShift/canshift-core | verified |
 <!-- packs:end -->

docs/packs.md

@@ -140,7 +140,9 @@ site is reported is part of the contract (`reported`):
   findings are claimed and folds compiler errors that one edit elsewhere resolves into a
   finding at that edit, as their evidence. The typescript pack folds the "Cannot find
   name" errors of the 6.0 `types: []` default into one at the tsconfig that leaves
-  `types` unset.
+  `types` unset. `pack test` counts a ground-truth site listed as such evidence as
+  covered, at the site's own rule and severity; evidence the truth does not name is not
+  scored.
 - `decision`: the pull request lists it under "Decisions for you", with the files and lines.
 - `test-follow-up`: an assertion is updated only after the migrated code made that test
   fail, and the test then has to pass (zod's reworded default messages).

packages/core/src/packs/tooling/pack-test.test.ts

@@ -142,6 +142,15 @@ it('scores an anchored finding at its anchor; the call sites listed as evidence
   expect(predictedSites(report, aiPack).map((s) => `${s.file}:${s.line} ${s.rule}`)).toEqual([
     'packages/lib/hook.ts:6 tool-context',
   ]);
+  // Evidence the ground truth names is covered by the anchor, at the truth's rule and severity;
+  // evidence it does not name (x.tsx:11) stays unscored.
+  const expected = [
+    { file: 'apps/b/src/y.tsx', line: 59, rule: 'generic', severity: 'breaking' as const },
+  ];
+  expect(predictedSites(report, aiPack, expected)).toEqual([
+    { file: 'apps/b/src/y.tsx', line: 59, rule: 'generic', severity: 'breaking' },
+    { file: 'packages/lib/hook.ts', line: 6, rule: 'tool-context', severity: 'breaking' },
+  ]);
 });
 
 it('counts a required move as moved, and a move check reports only a new major makes as explained', () => {

packages/core/src/packs/tooling/pack-test.ts

@@ -141,8 +141,14 @@ const key = (s: { file: string; line: number }): string => `${s.file}:${s.line}`
 /**
  * The sites `check` shows for the package, from its plan: the findings worth acting on,
  * grouped under the rule `fix` would apply. Repository-relative, as ground truth is written.
+ * An `expected` site an anchored finding lists as evidence is shown too (see below).
  */
-export function predictedSites(report: CheckReport, pack: Pack): ScoredSite[] {
+export function predictedSites(
+  report: CheckReport,
+  pack: Pack,
+  expected: ScoredSite[] = [],
+): ScoredSite[] {
+  const wanted = new Map(expected.map((e) => [key(e), e]));
   const own = new Set([...pack.rules.map((r) => r.id), ...pack.behavior.map((b) => b.id)]);
   const sites = new Map<string, ScoredSite>();
   for (const p of report.packages) {
@@ -184,6 +190,14 @@ export function predictedSites(report: CheckReport, pack: Pack): ScoredSite[] {
         // One site, one rule: the more severe group wins, as the plan lists it first.
         if (!sites.has(key(site))) sites.set(key(site), site);
       }
+    // Evidence the ground truth names is covered: `check` reports the line under its anchor
+    // (TypeScript 6's `types` errors under one tsconfig finding, #136's option cascades), so
+    // it scores as the site the truth expects, at that site's rule and severity. Evidence the
+    // truth does not name stays unscored.
+    for (const k of evidence) {
+      const site = wanted.get(k);
+      if (site && !sites.has(k)) sites.set(k, { ...site });
+    }
   }
   return [...sites.values()].sort((a, b) => a.file.localeCompare(b.file) || a.line - b.line);
 }
@@ -350,7 +364,7 @@ async function scoreRepo(
     if (installed.length > 0 && !installed.includes(entry.from))
       base.installedMismatch = `the lockfile has ${installed.join(', ')}, the ground truth says ${entry.from}`;
     const own = report.packages.filter((p) => p.name === pack.name);
-    const predicted = predictedSites(report, pack);
+    const predicted = predictedSites(report, pack, expected);
     return {
       ...base,
       ...companionOutcome(own, entry.with, (name) => dependsOn(project, dir, name, pack.name)),

packages/core/src/packs/typescript/RESEARCH.md

@@ -217,6 +217,13 @@ The recall figure measures the original truth's convention (every error line a s
 the engine's (one site per edit). It is not a count of breaks `check` misses: every folded
 error is listed under its tsconfig finding.
 
+**Scoring the folded evidence.** Since then `pack test` counts a ground-truth site that an
+anchored finding lists as evidence as covered by that finding, at the site's own rule and
+severity (`predictedSites`, as for #136's option cascades). The ground truth did not change.
+Re-scored 2026-10-15: breaking recall **99% (1,379 of 1,399)**, 0 breaking false positives,
+the same as before the fold. The 20 breaking misses are the second-round and same-change lines
+listed under "Measured `pack test`".
+
 ## Verified under UPT-48
 
 `pnpm uptide pack test typescript --write` on main (`530f90c`, UPT-48's gate), with its own
@@ -226,7 +233,7 @@ is unchanged (same digest as the candidate record).
 - Repositories: 7 public (gate: ≥2).
 - Real breaking sites: 1,399 expected (gate: ≥5).
 - Breaking false positives: 0 (gate: 0). Breaking precision 100%, recall 17% (240 of 1,399; the
-  folded `types` errors above).
+  folded `types` errors above), 99% (1,379) once the scorer counts folded evidence (above).
 - Behavior notes: precision 20%, recall 50% (2 TP, 8 FP, 2 FN), all `changed-defaults` anchors at
   `tsconfig*.json:1`. Notes do not count against a breaking-class gate.
 - Per-rule real-world true positives: `base-url` 4, `module-resolution-node10` 2, `target-es5` 1,

packages/core/src/packs/typescript/verification.json

@@ -4,10 +4,10 @@
   "repos": 7,
   "breaking": {
     "precision": 1,
-    "recall": 0.1716,
+    "recall": 0.9857,
     "falsePositives": 0,
-    "truePositives": 240,
-    "falseNegatives": 1159
+    "truePositives": 1379,
+    "falseNegatives": 20
   },
   "behavior": {
     "precision": 0.2,