test(eval): field-pattern regression fixture with xfail-tracked coverage gaps (6F.2)

[officialCodeWork]

Summary

Closes the fixture-vs-reality blind spot the 2026-07-15 field validation exposed: every gate scored 1.000 while the real-codebase run got 28% instance resolution, 0 RTK data sources, and 0 route nodes. This fixture (eval/fixtures/field-patterns) replicates those exact shapes:

• Two-hop barrel + rename (components/index.ts → ui/index.ts → definition) and tsconfig-path aliases (@ui, @store/*)
• Loadable(lazy(() => import())) page elements
• Route arrays in a separate file, spread-composed, passed to createBrowserRouter as an imported identifier
• RTK Query store — createApi base + two injectEndpoints slices (string-form and object-form query, a builder.mutation)
• Tests through a custom renderWithProviders wrapper

Mechanism

The golden asserts target behavior. 9 checks the scanner can't satisfy yet carry expectedFail naming the enabling step — 6F.3 (DataGrid instances + blast radius), 6F.4 (endpoint attributions), 6F.5 (routes, effect, journey). The harness's unexpected-pass gate removes stale marks the moment a capability lands.

checks.ts change: xfail-marked attributions stay out of the lineage precision/recall tallies while marked — a known-missing capability shouldn't depress global metrics.

Confirmed by the scan

• The @ui alias import yields no instance node at all for UsersPage's <Grid/> — matching the field's "72% never linked" diagnosis
• 0 route / data-source / hook nodes
• Already working (kept as passing checks): relative two-hop barrel rename, covered-by through the custom wrapper — so 6F.6 must find the actual field-breaking coverage variant

Verification

• pnpm eval: 280 pass · 0 fail · 9 xfail · 0 unexpected-pass, gate OK, all metrics 1.000

🤖 Generated with Claude Code