Polish follow-ups from the target-aging review (#280)

Non-blocking polish items from the adversarial review of #280 (all verified non-gating there):

1. Inertness regression guard: test_age_targets_defaults_off_leaves_surface_unchanged compares two off-paths; strengthen it to compare against a frozen pre-#280 registry hash so "byte-identical to the un-aged surface" is enforced against drift, not tautology.
2. Full ratio provenance: aging_factor_source records the numerator (build-year) CBO fact id only; add the denominator (source-year) fact id so the factor is reconstructable without inference (target_aging.py ~200–208, 316–334).
3. Symmetry nit: apply the 1900–2100 sanity bound to the string branch of _period_year (int branch has it).
4. Future: program benefit totals (SS/SNAP/TANF/SSI/Medicare premium) fall back to AGI growth if ever cross-period; a benefits/COLA index would be more faithful. Currently a documented no-op (all current-period).

Start with test_age_targets_defaults_off_leaves_surface_unchanged and the cited sections of target_aging.py around lines 200–208 and 316–334. Verify the frozen pre-#280 registry comparison, complete aging_factor_source provenance, and apply the year bound to the string branch of _period_year; the benefits/COLA item remains a documented no-op for now.