fix(dashboard): count rank summary per keyword, not per device

[shuvamk]

Problem

getRankSummary builds the dashboard rank scorecard. It counts the denominator per keyword but the numerators per device:

summary.trackedKeywords += result.rows.length;        // one row = one keyword
for (const row of result.rows) {
  for (const device of ["desktop", "mobile"] as const) {  // both devices
    const { position, previousPosition } = row[device];
    if (position !== null && position <= 10) summary.top10 += 1;
    ...
  }
}

Each row carries both a desktop and a mobile result for the same keyword, so a keyword ranking in the top 10 on both devices adds 2 to top10 but 1 to trackedKeywords. The summary can therefore report more top-10 keywords than tracked keywords (e.g. "2 in top 10 out of 1 tracked"). improved/declined have the same cross-basis inflation.

The tested sibling computeScorecards computes these per keyword for a single device (it takes a device param), confirming the intended unit is per-keyword — not both devices summed into a per-keyword denominator.

Fix

Extract a pure summarizeRankRows helper that classifies each keyword once, using its best (lowest) position across devices, and use it in getRankSummary. All four numbers now share the per-keyword basis, so top10/improved/declined can never exceed trackedKeywords.

Semantic note for reviewers: I collapsed the two devices to the keyword's best position — the natural reading for an overview scorecard ("how many of my keywords rank in the top 10"). The existing null-handling is preserved (a keyword only counts toward improved/declined when it has both a current and a previous position). If you'd prefer a different aggregation (e.g. a fixed primary device, or counting "new"/"lost" like computeScorecards does), the helper is the single place to change it.

Tests

New rankSummary.test.ts: a both-device top-10 keyword counted once; top10 never exceeding trackedKeywords; best-device position used for the top-10 check; a null-current keyword ignored; empty input zeroed.

Verification

• pnpm test — 715 passed (88 files)
• pnpm ci:check — prettier, knip, tsc --noEmit (×2), and oxlint all green

🤖 Generated with Claude Code

[sorcerai]

Reviewed exact current diff and tests. Summary counts each keyword once using its best device position, restoring top10 <= trackedKeywords.