Belgian selection: country-specific L0 from the full base (dense-informed init) once the BE target surface productionizes

The problem

populace-be currently inherits the US frozen 57,240-record selection (the certified US sparse support) as its record set, reweighted to Belgian targets. That inheritance is a convenience, not a choice: the US selection's value is accumulated optimization against the US target surface (populace#299's Build G head-to-head: frozen beats cold L0 decisively — but only because the frozen set embodies search against THAT objective). For Belgium the objective is different, so:

1. Records that would carry Belgian structure but didn't make the US cut are unavailable to Belgian calibration — the US selection silently binds BE.
2. There is no Belgian incumbent lineage to warm-start from; Belgium's first real selection must be fresh.

At today's 21-marginal-target surface this doesn't bite (21/21 within 1.8%, healthy ESS). It will bite as #263 builds out the real Belgian surface.

The ask

When the Belgian target surface productionizes (#263), run Belgium's own selection:

• Pool: the full 337,704-record base (not the inherited US 57k).
• Objective: the productionized Belgian registry.
• Form: cold in lineage, warm in initialization — run the Belgian DENSE solve first and initialize L0 selection probabilities proportional to dense weights (the dense-informed init from populace#346), plus criticality weighting on Belgium's protected targets. This avoids the thin-tail failure mode that killed cold L0 in the US head-to-head (populace#299) without pretending a US-optimized lineage transfers across countries.
• Output: the first entry of a Belgian selection lineage, recorded via the #330 selection-manifest machinery so future BE releases warm-start per #346.

Sequencing: with or immediately after #263; before any certified populace-be release channel (#265). Prereq alignment: the transfer-paper input-quality work (fx conversion done; missing components per #262/#263) should land first — selection against a surface computed from wrong inputs bakes in wrong choices.

Refs: #259 (EPIC), #263, #265, #346 (informed-init design), #330 (selection-manifest seam), #299 (frozen-vs-cold evidence).

Start with #263 for the production Belgian target surface, then read #346 for dense-informed initialization and #330 for selection manifests. The work should select from the full 337,704-record base using the Belgian registry, protected-target criticality weighting, and dense initialization. It is done when the first Belgian selection lineage entry is recorded before the certified release channel.