[quality] e2e region union does not fold single-line '||' fallback arms across the real and variant builds, so correct variant tests lower the gate #1066

Finding

The data-variant build is the project's documented mechanism for covering
branches that turn on a document-level field
(tests/tools/e2e-data-fixtures.cjs, tests/e2e/data-variants.spec.js). For
one shape of branch it does not work: the variant build executes the arm, and
the report still counts it uncovered. Writing the test correctly makes the
gate number go down.

What happens

getRegionCoverage() (tests/tools/e2e-coverage-report.mjs:316-345) keys each
region on original-source coordinates and unions across scripts:

const key = [
  line,
  location.start?.column ?? 0,
  location.end?.line ?? line,
  location.end?.column ?? 0,
].join(':');

The real build and the variant build are two different minified chunks. Their
block boundaries carry no mapping at most endpoints, so an endpoint snaps to
whatever mapping precedes it — and the two chunks snap differently. The same
branch therefore produces two keys that never fold.

Dumping every region key for src/components/MetricsDashboard/index.js over a
local run that visits both sites (88be21cb = real build, caf361b6 =
variant):

65:21:65:31  count=0   39a40737.88be21cb.js     <-- reported uncovered
65:21:65:38  count=1   39a40737.caf361b6.js     <-- same arm, variant ran it
76:35:76:49  count=0   39a40737.88be21cb.js
76:35:76:53  count=1   39a40737.caf361b6.js
85:21:85:33  count=0   39a40737.88be21cb.js
85:21:85:40  count=1   39a40737.caf361b6.js

Same line, same start column, end column differs by 5–7. These are the
fallback arms of

• (lifecycle?.cards || []) — line 65
• (lifecycle?.trends || {}) — line 76
• (lifecycle?.omitted || []) — line 85

and the guard is demonstrably present and executed in the variant chunk:

lifecycleGrid_j1K8",children:(e?.cards||[]).map(...)

Why #1051 does not already cover this

#1051 folds out phantom regions,
but only multi-line zero-count regions — single-line ones are left alone on
purpose, because several can share a line and the per-line fold cannot tell
them apart. Every drifted half above is single-line, so isPhantomRegion
cannot touch it. Re-rendering the same run with #1051 applied leaves lines
65 76 85 124 174 uncovered exactly as before.

This is also why the existing variant tests look like they work: AwardsTimeline
and ReferenceArchitectures hang off ternaries, whose drifted halves are
multi-line and therefore are folded. The same dump for
ReferenceArchitectures:

22:6:23:11   count=0   a1931b28.cbfec487.js    <-- multi-line, #1051 folds it
22:6:34:18   count=1   a1931b28.9e6c34b7.js

So the mechanism silently works for a ? b : c and silently fails for
a || b.

Why it matters: a correct test makes the gate worse

I wrote the variant overlay and three specs for the five absent-data fallbacks
in MetricsDashboard (referenceArchitectureLifecycle, series,
breakdowns cleared in tests/e2e/fixtures/data-variants/metrics.json). All
301 specs pass and the variant page provably renders the fallback arms. The
report's response:

Visiting the variant route admits that chunk's regions into the union for the
first time: 12 become covered, 23 enter the denominator, and the five targeted
arms stay uncovered. .github/workflows/ci.yml gates on
--check-source-regions 80, so the correct test fails the gate.

That is not a MetricsDashboard problem. Every remaining uncovered e2e region
outside the open PRs is a document-level case that needs the variant build —
RadarReports lines 6 and 24, GroupLinkStatus lines 7, 23, 32 and 33 — so
this blocks the rest of the e2e backlog, and currently penalises anyone who
attempts it.

Evidence and provenance

• Unit: npm run test:unit:coverage (TZ=UTC node tests/tools/coverage-report.mjs,
  node v26.10.0), local, revision 900592b — src files 100.00% lines /
  99.84% regions; src/components/MetricsDashboard/index.js 100.00% / 100.00%.
  The gaps below are e2e-only.
• E2E: local npm run build:e2e:coverage then npm run test:e2e:coverage
  with E2E_COVERAGE_DIR / E2E_COVERAGE_RUN_ID set, revision 900592b,
  298 passed (301 with the experimental specs). This reproduces CI job
  End-to-end coverage, workflow Validate repository, run
  37164361552
  attempt 1, to within one region: CI 468 / 379 / 80.98%, local 465 / 377 /
  81.08%.
• Region keys dumped by instrumenting the union site in
  tests/tools/e2e-coverage-report.mjs:530-536 with the script URL.

Recommendation

Make region identity survive the same source branch being compiled into two
chunks, rather than widening the phantom fold:

1. Normalise the region key so a drifted endpoint cannot split one branch.
   Keying on startLine:startColumn plus the branch's ordinal within that
   line, and taking the maximum count over the group, folds every pair above
   while still distinguishing the several single-line arms that share a line —
   which is the property test: fold out phantom e2e coverage regions over covered lines #1051 had to preserve.
2. Add a regression test to tests/e2e-coverage-report.test.mjs driving the
   union with two synthetic scripts whose maps give one branch the same start
   and different end columns, asserting it folds to a single region counted
   covered when either script hit it.
3. Re-derive --check-source-regions afterwards. The floor is set in
   .github/workflows/ci.yml.

Coordination

• The fix lands in getRegionCoverage() / the union in
  tests/tools/e2e-coverage-report.mjs, which open PR
  #1051 is currently rewriting.
  This issue should not be implemented until test: fold out phantom e2e coverage regions over covered lines #1051 lands, and should then
  build on it rather than replacing it — the phantom fold and the key
  normalisation are complementary, not alternatives. I have deliberately not
  opened a competing PR.
• Step 3 touches .github/workflows/ci.yml. That path cannot be pushed by this
  lane's token, so re-deriving the floor needs a human or an agent with the
  Workflows permission.

Completion criteria

• Region identity folds one source branch across the real and variant chunks when only an end column differs
• A regression test in tests/e2e-coverage-report.test.mjs covers the differing-end-column union
• MetricsDashboard lines 65, 76, 85, 124 and 174 can be retired by a variant test without lowering src files regions
• --check-source-regions re-derived from the corrected number

Priority

• Impact: high (the merge gate penalises correct tests, and this blocks every remaining document-level e2e gap)
• Effort: medium

🐝 Hive Agent: quality | Instance: hosted-available-lke648397-260827-5n31 | SHA: 900592b

— hive: agent=quality backend=copilot model=claude-opus-5 copilot=1.0.88