fix: reject partial hourly drills

[NikolayS]

What changed

• preserves exact unfiltered/database-only partial-hour aas and timeline reads through rollup_1h.minute_counts
• requires hour-aligned bounds when a reader must use hourly wait/query arrays
• rejects explicitly requested sub-hour timeline and chart buckets on rollup_1h
• widens omitted buckets to the safe one-hour grain automatically
• documents the retained-precision boundary in README and catalog comments

Why

An hourly dimensional row represents a full hour. The readers previously selected whole rows by their start timestamp and divided their complete counts by a clipped partial window. A single-backend half-hour IO read could therefore report AAS 2.0, while a cross-hour read could omit the overlapping first hour.

The current storage format cannot reconstruct partial-hour wait/query dimensions exactly. Rejecting unsupported shapes is safer than fabricating precision.

Impact

The data API now raises an actionable error for unsupported partial-hour dimensional reads. Unfiltered minute-total reads remain exact and unchanged.

Validation

Postgres 18.3:

• RED at 5da0e13: filtered half-hour aas returned instead of rejecting
• GREEN at 1aa6bf3:
  • unfiltered half-hour remains avg_aas = 1.00, backend_seconds = 1800
  • filtered aas, top, dimensional compare, explicit sub-hour timeline, and explicit sub-hour chart reject
  • aligned filtered hour remains exact at avg_aas = 1.00, backend_seconds = 3600
  • omitted timeline/chart buckets auto-select two one-hour rows
• existing 2.0 rollup-source reader tests pass
• existing rollup_1h minute-seam tests pass
• existing calendar-bucket and spike-first ordering tests pass
• installer re-apply with ON_ERROR_STOP=1 passes
• git diff --check main...HEAD passes

Closes #130

[NikolayS]

Superseded by #182, merged as a1f0a30. #182 fixes the same #130 partial-hour failure across the full reader surface by snapping hour-only plans outward and disclosing effective bounds/buckets, while preserving exact minute_counts-backed database/unfiltered reads and zero-argument/default calls. Closing this narrower hard-error implementation in favor of the merged successor.