fix: avoid stale rollup reads

[NikolayS]

What changed

• adds a behavioral regression for a greater-than-one-hour window where raw samples cover the requested window but rollup_1m stops after its first minute
• makes ash._pick_source_agg() require the rollup watermark to reach until when raw covers since
• preserves the existing partial-rollup fallback when raw cannot answer the window start

Why

The 2.0 aggregate readers selected rollup_1m from its retention start alone. If the rollup worker lagged or stopped, the readers returned plausible but incomplete values and omitted newer raw load.

This is a release-candidate SQL correction on a release/* branch because the beta installer is already stamped under sql/.

Impact

The correction applies to the new aggregate surface that delegates to _pick_source_agg(): aas, timeline, periods, non-tie top, and chart.

Validation

• RED on Postgres 18 at 54362ff: stale rollup must fall back to raw, got rollup_1m
• GREEN on Postgres 18 at 6123780: exact source, covered-minute, and backend-second assertions pass
• fresh install and full installer re-apply with ON_ERROR_STOP=1 pass
• git diff --check main...HEAD passes

Closes #122

[NikolayS]

Superseded by #199, which keeps stale wide rollup reads honest without losing the healthy rounded-up dashboard fast path and landed on current main as 9efad9c. Closing this earlier attempt in favor of that successor; issue #122 remains open for the explicitly unclaimed older-than-physical-raw completeness case.