Conserve impact for background production-edge TDs in traverse_background

[TimoDiepers]

Stacked on top of #197 (base: `feat/premise-temporal`) so it can be reviewed in isolation.

Problem

With `traverse_background=True`, when the variant-aware descent reaches a background node that carries a production-edge temporal distribution, the production TD was applied only to the node's child expansion, never to the edge that produces the node. Two symptoms follow:

• KeyError in get_time_mapping_key — the node is consumed at cohort-years it was never registered as a producer at (producer/consumer year bands diverge).
• N× over-count — _join_datetime_and_timedelta_distributions tiles the producer TD's amounts and drops the consumer-side cohort weights, so each spread cohort emits the full exchange coefficient instead of its weighted share. A 3-cohort production TD inflates the score ~3×.

Fix

Fold the producer's own production-edge TD into the effective producer TD of the edge at both background descent sites (_emit_variant_split_for_consumer_date and _descend_variant_subtree), via a new _fold_production_td helper that takes the outer product of amounts and outer sum of dates (dates sum, weights multiply). The node is then registered at exactly the production-TD-weighted cohorts it is consumed at:

• producer band == consumer band by construction → no more KeyError;
• each cohort carries exchange_weight × production_weight → impact conserves.

This mirrors the existing FU-seed cohort-split pattern in build_edge_timeline. The change is confined to VariantBackgroundMixin (background proxy-descent) — the foreground/explicit product-process path, the matrix traversal, and _join_datetime_and_timedelta_distributions are untouched.

As defense-in-depth for a separate, not-yet-minimally-reproduced premise dual-path case, get_time_mapping_key also snaps a background consumer lookup to the nearest registered year of the same node (and now logs a warning when it does) instead of raising.

Tests

New `tests/test_background_production_td.py` (both `priority` and `bfs` engines), each asserting time-explicit `static_score == base_lca.score`:

• first-level descended production-TD node conserves;
• deeper descended production-TD node conserves;
• convergent production-TD node (reached via two parents) conserves — verified non-vacuous (fails on pre-fix code).

Full suite green.