[needs visual review] #1138 FinOps index/object heatmap redesign (both phases, both apps)

[erikdarlingdata]

Implements #1138 — the FinOps index/object heatmap redesign, both phases, both apps (Dashboard + Lite), on existing collected data. Follows the 3-review-round plan in plans/finops-index-heatmap-redesign.md.

This PR is code-complete but needs your visual sign-off — I cannot launch the apps from the worktree (single-instance would evict your running Dashboard/Lite and kill the MCP servers), so the heatmap rendering, cell shading, and drill navigation need an eyes-on pass. The code itself is done: both apps build green, tests pass, and all new T-SQL was validated live against a real PerformanceMonitor database (SQL2025).

Phase 1 — Object-growth heatmap + Storage Growth drill

Replaces the two flat tabs Object Sizes & Growth (#5, removed) and Index Usage (#6, folded) with a drill rooted at Storage Growth (#4):

• Storage Growth grid → double-click a database row (or the new "Show objects" context-menu item) → that database's object-growth heatmap: rows = top-N objects by reserved-MB growth over the window, X = daily buckets, color = absolute reserved MB (reuses the query-heatmap renderer's 0→NaN behavior unchanged, so a blank cell = object absent/empty that day). A companion per-object grid sits below it.
• Double-click an object → its per-index detail (size + the old Index Usage seeks/scans/lookups/updates).
• Back/breadcrumb navigation; a window selector (7/30/90d, default 30); an in-UI substrate caveat ("object / reserved footprint … may not explain file-level growth").
• Producer is two-step + DB-scoped for perf (plan §3A): rank the top-N cheaply via the 2-snapshot diff, then series-scan only those — never the uncovered all-objects 90-day scan that caused [BUG] index_object_stats collector times out (30s) as a single all‑or‑nothing cross‑database sweep - yields no index data #1135. Rollup key is name-based (database, schema, table) (plan K1).
• The standalone Object Sizes / Index Usage MCP read methods are untouched (kept for the MCP tools).

Phase 2 — Locking color-scaled wait-type grid + drill

Turns Locking & Contention (#7) into a heatmap-in-grid (plan §3B — not a ScottPlot heatmap, so no M1/M2 issues):

• The four *_wait_in_ms cells are shaded per-column (log scale, max=0 guarded) by their wait category's hue — row/page lock = Lock, page latch = Buffer Latch, page-io-latch = Buffer IO — via a shared themed value→brush converter. The numeric values stay visible in every cell.
• A database selector (per-database latest, so "all databases" shows each DB's newest row); double-click an index → full lock/latch counters + identity, with back.
• Read layer gains an optional DB filter, a per-DB-latest CTE, and the two latch *_count columns; a new GetIndexLockingDatabasesAsync feeds the selector.
• Doc-drift fix (plan R7): install/02, install/55, and the Lite collector header all claimed lock-wait deltas are "computed in the read layer" — no such logic exists and M1 shows it wouldn't be safe; corrected to "raw cumulative totals, no delta".

Shared / parity

To keep the two hand-duplicated apps from drifting, the parity-critical logic is shared once:

• PerformanceMonitor.Common/FinOpsHeatmap.cs — pure ranking, long→matrix pivot, per-column log color-scale.
• PerformanceMonitor.Ui/FinOpsHeatmapRenderer.cs — the query-heatmap renderer with count couplings stripped (MB colorbar, M/d axis).
• PerformanceMonitor.Ui/HeatIntensityToBrushConverter.cs — themed cell shading.

Tests

• Pure unit tests in both Dashboard.Tests + Lite.Tests: top-N ranking, the long→matrix pivot (incl. missing-cell=0 + ordering), and the per-column color-scale math (incl. the max=0 guard).
• Lite integration tests (DuckDB): the object-growth producer (growth ranking + daily series), the per-index detail, and the locking DB selector/filter.
• Both apps build green; new T-SQL validated live on SQL2025.

Reviewer-driven fixes already folded in

• Deterministic ORDER BY tiebreaker on the growth ranking so Lite's two queries (and Dashboard vs Lite) always pick the same top-N set + order; the heatmap rows and companion grid now share one ranking authority.
• T-SQL polish on the producer (COUNT_BIG, unique-clustered #ranked, TABLOCK, OPTION hints, CONVERT over CAST).

Not in this PR

No merge — the only thing gating merge is your visual sign-off. No newly collected columns (read-layer + UI only). The other adjacent tabs (#8 Database Sizes is per-file, #9/#10/#11) are unchanged per the plan's scope.

🤖 Generated with Claude Code