smart-nav-menu: one pinned-list resolver, not two (DEFECTS #18)

[roncodes]

Fixes DEFECTS #18 — the last known blocker to a stable 100% gate.

The symptom

The suite's branch total varied by ±1 between identical runs. All 5,196 tests passing, identical statement totals, different branch counts. A hard 100% gate flaps on that with no code change, exactly as #4 did.

Diagnosis — the artifact named it in one step

Rather than reasoning about mechanism, I captured coverage-final.json from two runs that disagreed and diffed them per-file, per-branch:

addon/components/layout/header/smart-nav-menu.js
  branches 82 vs 83
  branch id 21 path 1 @ line 344: runA=0 runB=1 (if)

That is if (item) pinned.push(item) — the "a pinned id matches no item" path.

The cause is duplication, not flakiness

The same "resolve pinned IDs in the user's saved order, dropping stale ones" loop existed in two methods:

• _distributeFromAllItems() runs on render, and the existing test "stale ids in the saved list are skipped" covers its copy deterministically.
• _recalculate() carried its own copy, reached only when a ResizeObserver fires through scheduleOnce('afterRender') — browser layout timing, which no test can force.

So the branch was genuinely exercised, in two places, one of them a coin flip.

The fix

Extracted _pinnedItems(pinnedIds, allItems); both callers use it. One branch where there were two, covered by the test that already existed.

This is worth doing on its own merits. Two copies of stale-ID handling is the shape where someone fixes one and leaves the other quietly wrong. The stable branch count is a consequence of removing the duplication, not the point of it.

Verification

run C: 5196 pass, 0 fail, branches 5908/6332
run D: 5196 pass, 0 fail, branches 5908/6332
C vs D per-file diff: no differences at all

Not merely matching totals — the diff compares every file's counts and then every individual branch id, so nothing can cancel out. Lint exits 0.

The branch total drops 6334 → 6332, which is exactly the two paths of the removed duplicate.

Note on method

#4 — the previous nondeterminism — cost two wrong diagnoses because I reasoned about the mechanism first and was confidently wrong twice, once applying a "fix" that made it strictly worse. #18's entry was written to say: diff two disagreeing artifacts and let them name the file before forming a theory. That is what happened here, and it took one step.