feat(gain): colony gain drift + savings_drift_report MCP tool

[NagyVikt]

Summary

Adds a long-run regression detector that flags tools whose median tokens-per-call has drifted up or down between a baseline window and a recent window. Pure read path against existing `mcp_metrics` — no schema change.

• CLI: `colony gain drift [--baseline-days 14] [--recent-days 3] [--min-calls 20] [--threshold 1.25] [--down-threshold 0.75]`
• MCP: `savings_drift_report` (mirrors `savings_report` shape)
• Storage: `mcpTokenDriftPerOperation()` + `mcpMetricsMinTs()`
• Classifier: `up_drift` / `down_drift` / `new_tool` / `gone` / `insufficient_data` / `stable`

Closes `⏳ Long-run regression detector that flags when a tool's median tokens-per-call drifts up` under README §v0.x "Receipts and observability".

OpenSpec

`openspec/changes/gain-drift-detector-2026-05-16/CHANGE.md`

Design adaptation

The original Plan used a correlated `LIMIT 1 OFFSET (COUNT(*)-1)/2` for the per-operation median. SQLite forbids outer aggregate refs in scalar-subquery `OFFSET`. Switched to `ROW_NUMBER() OVER (PARTITION BY operation ORDER BY tpc)` with a CTE join — same semantics, cleaner, supported by bundled better-sqlite3 (SQLite 3.49.2 verified).

Test plan

• `pnpm --filter @colony/storage test` — 165 pass (18 mcp-metrics, +7 new)
• `pnpm --filter colonyq typecheck` — clean
• `pnpm --filter colonyq test -- gain-drift` — 5 pass
• `pnpm --filter @colony/mcp-server test` — 299 pass
• `pnpm --filter colonyq build` — clean

Merge order

Touches `packages/storage/src/storage.ts`. If #573 and #574 (scenarios) merge first, this should be third. The coach-mode PR also extends `storage.ts` independently — whichever lands second needs a trivial rebase.

🤖 Generated with Claude Code