[CmdPal] Replace main-page magic-number scoring with a principled tiered ranker

[michaeljolley]

Summary

Users report Command Palette main/root page results are wrongly ranked: the thing they want exists but is buried below junk. Today's ordering is a flat weighted sum of hand-tuned magic numbers in MainListPage.ScoreTopLevelItem (alias +9001, match x10, subtitle -4/2, extension /1.5, fallback x0.5) plus an additive frecency weight, which makes obvious matches lose to noise and is effectively untestable.

This PR re-architects the main-page scoring into a principled, testable, multi-signal ranker. It is the first atomic step of a larger search-quality effort (see follow-ups below).

What changed

• New MainListRanker (Microsoft.CmdPal.UI.ViewModels/Commands/MainListRanker.cs):
  • A hard tier ladder decides ordering: alias-exact > exact-title > prefix > acronym/word-boundary > fuzzy > fallback floor.
  • A normalized within-tier score (lexical quality + frecency + alias-substring nudge + provider bonus) only reorders items that already share a tier, so a weak signal can never jump an item over a better lexical match.
  • Tier + within-tier score are packed into a single sortable int, so the existing score-descending sort (InternalListHelpers.FilterListWithScores) is unchanged.
  • Defines IRankSignal / IRerankStage seam interfaces so a future on-device semantic or small-LLM reranker can slot in later as a post-first-paint pass. No model ships here.
• ScoreTopLevelItem refactor: the magic-number tail is replaced by tier classification + the within-tier score. The within-tier math deliberately mirrors the old balance, so items that share a tier keep their long-standing relative order; only cross-tier behavior changes.
• Tests: updated the ranking invariants to assert the new contract - heavy usage reorders within a tier but never crosses a tier boundary.

Scope / safety

• Host-side only. The extension SDK toolkit (ListHelpers / FuzzyStringMatcher) and extension-owned list pages are untouched, keeping this atomic and free of ABI/extension-compat risk.
• Deterministic; no new dependencies; no settings schema or IPC changes.

Validation

• Build CommandPalette.slnf (x64 / Debug): exit 0.
• All 146 Microsoft.CmdPal.UI.ViewModels.UnitTests pass.

Follow-ups (separate PRs)

These were intentionally deferred to keep this PR atomic and to give privacy/schema-sensitive work its own review:

1. frecency-redesign - timestamped, time-decayed frecency (exp decay + log(uses)), larger store, migration of persisted history; feeds the within-tier signal.
2. provider-weighting - per-provider Lower/Normal/Higher control in Settings (schema + XAML), applied as a within-tier bonus.
3. fast-first-paint - render deterministic command/app results instantly and slot slow out-of-proc fallback results in asynchronously.
4. relevance-harness - golden-set MSTest harness asserting expected top-N ordering.
5. search-telemetry - opt-in, privacy-tagged search telemetry (never logs raw query text).

Co-authored-by: Copilot App 223556219+Copilot@users.noreply.github.com