Evaluation harness: support aged targets (declared fact-ratio levels) in precomputed checkpoints

Found publishing Microcosm-BE v0.5 (2025 simulation) to the cross-dataset page. v0.5's release surface contains 233 aged targets — 2025 levels derived as published fact × declared FPB fact-ratio with both record ids and the factor in the manifest aging block. The precomputed-checkpoint adapter (#164) enforces two constraints that are jointly unsatisfiable for such releases:

1. every checkpoint row's period must equal its Chronicle facts' period (precomputed checkpoint row 'statbel_b1_decile_01_total_tax' period does not match its Chronicle facts — the row is a 2025 level over …income_year2023… facts);
2. the checkpoint row set must cover the authoritative Microcosm release surface (filtering the aged rows out fails with differs from the authoritative Microcosm release surface; missing=[233 rows]).

The US path solves this with first-class aging transforms (microcosm_aging, period treatment aligned_fact). The BE precomputed path needs the equivalent: accept a per-row alignment declaration (numerator/denominator fact ids + factor, mirroring the release manifest's aging block), verify the arithmetic against the named facts, and surface it as the existing aligned_fact period treatment with calibration_exposure intact — never silently accepting a mismatched period.

Until then the datasets page stays on the v0.4 run (evaluation-f28ca06a…, unaged surface) while the calibration-fit page serves v0.5; the v0.5 checkpoints are in population-rerun/out/v05b/evaluation/ ready to re-run once supported.

Start with the precomputed-checkpoint adapter referenced in #164 and compare it with the US microcosm_aging path. Inspect the release manifest’s aging block and the checkpoints in population-rerun/out/v05b/evaluation/. Done when the v0.5 checkpoints can be rerun with verified aligned facts, preserved calibration_exposure, and no silent period mismatch.