feat(evals): parent-computed selection propagates to shard children

The sharded runner computed diff selection once, then each of its 48-73
children recomputed it at module load — including, on touchfiles-diff
branches, a per-child bun subprocess evaluating the old data file (20s
timeout each). The parent now serializes {version, selected, reason} as
EVALS_SELECTION_JSON into the shard env; e2e-helpers adopts it at load.
Fail-open preserved: any parse/shape violation → ONE stderr warning +
local recompute; absent env → silent local compute (non-sharded
entrypoints unchanged). Drift test pins parent→child round-trip to
identical selection decisions plus the malformed/absent cases.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

test/helpers/e2e-helpers.ts

@@ -61,7 +61,55 @@ export function computeDiffSelection(
   return selection.selected;
 }
 
-export let selectedTests: string[] | null = computeDiffSelection(E2E_TOUCHFILES, 'E2E'); // null = run all
+/**
+ * Parse the sharded paid runner's precomputed selection (EVALS_SELECTION_JSON,
+ * written by serializePaidDiffSelection in scripts/test-paid-shards.ts).
+ * Returns { selected: null } for run-all. THROWS on any parse/shape failure —
+ * resolveModuleSelection turns that into a fail-open local recompute.
+ */
+export function parseEvalsSelectionJson(raw: string): { selected: string[] | null; reason: string } {
+  const parsed: unknown = JSON.parse(raw);
+  if (!parsed || typeof parsed !== 'object' || Array.isArray(parsed)) throw new Error('not an object');
+  const { selected, reason } = parsed as { selected?: unknown; reason?: unknown };
+  if (selected !== null
+    && !(Array.isArray(selected) && selected.every((s) => typeof s === 'string'))) {
+    throw new Error('selected must be null or string[]');
+  }
+  return {
+    selected: selected as string[] | null,
+    reason: typeof reason === 'string' ? reason : 'parent selection',
+  };
+}
+
+/**
+ * Resolve the module-load E2E selection: prefer the parent shard runner's
+ * EVALS_SELECTION_JSON — skipping this module's own git walk and, when
+ * touchfiles-data.ts is in the diff, the per-child bun subprocess that
+ * evaluates the old data file (test-selection.ts map-diff path, one per
+ * shard). On ANY parse/shape failure, fall back to computing locally
+ * (fail-open preserved) with one stderr warning.
+ */
+export function resolveModuleSelection(
+  raw: string | undefined,
+  compute: () => string[] | null,
+  stderrWrite: (text: string) => void = (text) => process.stderr.write(text),
+): string[] | null {
+  if (raw) {
+    try {
+      const { selected, reason } = parseEvalsSelectionJson(raw);
+      stderrWrite(`\nE2E selection (parent-propagated: ${reason}): ${selected === null ? 'all' : selected.length} tests\n`);
+      return selected;
+    } catch (err) {
+      stderrWrite(`WARNING: malformed EVALS_SELECTION_JSON (${err instanceof Error ? err.message : String(err)}) — falling back to local selection\n`);
+    }
+  }
+  return compute();
+}
+
+export let selectedTests: string[] | null = resolveModuleSelection(
+  evalsEnabled ? process.env.EVALS_SELECTION_JSON : undefined,
+  () => computeDiffSelection(E2E_TOUCHFILES, 'E2E'),
+); // null = run all
 
 // EVALS_TIER: filter tests by tier after diff-based selection.
 // 'gate' = gate tests only (CI default — blocks merge)

test/paid-selection-propagation.test.ts

@@ -0,0 +1,117 @@
+/**
+ * Parent/child selection-drift pins for EVALS_SELECTION_JSON.
+ *
+ * The sharded paid runner computes the diff selection ONCE in the parent
+ * (computePaidDiffSelection in scripts/test-paid-shards.ts), serializes it
+ * (serializePaidDiffSelection) into every shard child's env, and
+ * test/helpers/e2e-helpers.ts adopts it at module load (parseEvalsSelectionJson
+ * via resolveModuleSelection) instead of re-deriving it per shard — which,
+ * whenever touchfiles-data.ts was in the diff, spawned one bun subprocess PER
+ * CHILD to evaluate the old data file (test-selection.ts map-diff path).
+ *
+ * These pins hold the two sides to IDENTICAL selection decisions across the
+ * serialize/parse boundary, and the child to fail-open (local recompute with
+ * one stderr warning) on any parse/shape failure.
+ */
+
+import { describe, test, expect } from 'bun:test';
+import {
+  computePaidDiffSelection,
+  serializePaidDiffSelection,
+  type PaidDiffSelection,
+} from '../scripts/test-paid-shards';
+import { parseEvalsSelectionJson, resolveModuleSelection } from './helpers/e2e-helpers';
+
+/** The parent's per-test decision shape (PaidDiffSelection.selectedNames). */
+const parentWouldRun = (selection: PaidDiffSelection, name: string): boolean =>
+  selection.selectedNames === null || selection.selectedNames.has(name);
+
+/** The child's per-test decision shape (testIfSelected / describeIfSelected). */
+const childWouldRun = (selected: string[] | null, name: string): boolean =>
+  selected === null || selected.includes(name);
+
+const NAMES = ['qa-workflow', 'review-army', 'ship-docsync', 'unmapped-test'];
+
+describe('EVALS_SELECTION_JSON parent -> child propagation', () => {
+  test('a concrete selection round-trips to identical decisions', () => {
+    const fixture: PaidDiffSelection = {
+      selectedNames: new Set(['qa-workflow', 'ship-docsync']),
+      reason: 'diff',
+      totalTests: 4,
+    };
+    const parsed = parseEvalsSelectionJson(serializePaidDiffSelection(fixture));
+    expect(parsed.selected).toEqual(['qa-workflow', 'ship-docsync']);
+    expect(parsed.reason).toBe('diff');
+    for (const name of NAMES) {
+      expect(childWouldRun(parsed.selected, name), name).toBe(parentWouldRun(fixture, name));
+    }
+  });
+
+  test('run-all (null) round-trips to null — child runs everything', () => {
+    // computePaidDiffSelection is the REAL parent function; EVALS_ALL is its
+    // git-free path, so the serializer sees input exactly as produced.
+    const selection = computePaidDiffSelection({ EVALS_ALL: '1' } as NodeJS.ProcessEnv);
+    expect(selection.selectedNames).toBeNull();
+    const parsed = parseEvalsSelectionJson(serializePaidDiffSelection(selection));
+    expect(parsed.selected).toBeNull();
+    for (const name of NAMES) {
+      expect(childWouldRun(parsed.selected, name)).toBe(parentWouldRun(selection, name));
+    }
+  });
+
+  test('empty selection stays empty — nothing selected is NOT run-all', () => {
+    const fixture: PaidDiffSelection = { selectedNames: new Set(), reason: 'diff', totalTests: 4 };
+    const parsed = parseEvalsSelectionJson(serializePaidDiffSelection(fixture));
+    expect(parsed.selected).toEqual([]);
+    for (const name of NAMES) {
+      expect(childWouldRun(parsed.selected, name)).toBe(false);
+      expect(parentWouldRun(fixture, name)).toBe(false);
+    }
+  });
+
+  test('parser THROWS on malformed JSON and wrong shapes', () => {
+    expect(() => parseEvalsSelectionJson('{"selected": ')).toThrow();
+    expect(() => parseEvalsSelectionJson('null')).toThrow();
+    expect(() => parseEvalsSelectionJson('[1,2]')).toThrow();
+    expect(() => parseEvalsSelectionJson('{"selected": 42}')).toThrow();
+    expect(() => parseEvalsSelectionJson('{"selected": ["a", 7]}')).toThrow();
+  });
+
+  test('malformed EVALS_SELECTION_JSON falls back to local compute with one stderr warning', () => {
+    const warnings: string[] = [];
+    let computed = 0;
+    const result = resolveModuleSelection(
+      '{"selected": 42}',
+      () => { computed += 1; return ['locally-computed']; },
+      (text) => warnings.push(text),
+    );
+    expect(result).toEqual(['locally-computed']); // fail-open preserved
+    expect(computed).toBe(1);
+    expect(warnings.length).toBe(1);
+    expect(warnings[0]).toContain('EVALS_SELECTION_JSON');
+  });
+
+  test('absent env var computes locally, silently (non-sharded entrypoints unchanged)', () => {
+    const writes: string[] = [];
+    let computed = 0;
+    const result = resolveModuleSelection(
+      undefined,
+      () => { computed += 1; return null; },
+      (text) => writes.push(text),
+    );
+    expect(result).toBeNull();
+    expect(computed).toBe(1);
+    expect(writes.length).toBe(0);
+  });
+
+  test('a valid env var short-circuits local derivation entirely', () => {
+    let computed = 0;
+    const result = resolveModuleSelection(
+      serializePaidDiffSelection({ selectedNames: new Set(['a']), reason: 'diff', totalTests: 1 }),
+      () => { computed += 1; return null; },
+      () => {},
+    );
+    expect(result).toEqual(['a']);
+    expect(computed).toBe(0); // no git walk, no map-diff bun subprocess
+  });
+});