base-m collapses partnership_self_employment_net_earnings to signed near-cancellation (+47.7B/−48.1B legs)

Found by the export-mass parity gate on Build M sparse attempt 10 (release id populace-us-2024-buildm-sparse-rmloss100-5482d38-20260715T114657Z): partnership_self_employment_net_earnings exported −$22.72B against the reference's +$61.74B (−136.8%, far beyond the ±50% band).

The gate is right that mass was lost, and the loss is in the base pool, not the solve:

base-m's column decomposes to a positive leg of +$47.71B against a negative leg of −$48.06B — near-perfect cancellation. The column is signed, sparse (4,467 nonzero of 865k persons, ~0.5%), and heavy-tailed on both sides: the worst case for leaf-averaged QRF imputation. base-j had already compressed it (+$61.7B → +$12.2B); the base-m rebuild (staged builder #406 chains, #415 leaf bound in effect) pushed the net to zero. The solve then amplifies in either direction because nothing pins the column — Build J's solve happened to drag it up in-band (−37.6%), Build M's dragged it down and out.

Reachability is NOT the issue (unlike the miscellaneous_income / #393 case): at ratio-5 the positive leg alone could reach $238B. The pool's signed near-cancellation is.

Remedy, in order:

1. Diagnose the QRF chain for this column in the staged base builder: does the near-discrete guard, the leaf bound, or donor-pool construction destroy the positive/negative separation? Compare per-leg distributions donor-PUF vs imputed.
2. Fix and rebuild base (per-target checkpointing makes this a single-chain rebuild, not a full base rebuild).
3. Longer term (architecture week): identify the column with a real target the way Build H did for estate_income — SE-earnings concepts can anchor to SOI Schedule SE statistics — so the solve stops treating it as a free dimension. Same applies to every unpinned column the parity gate watches.

A reviewed parity exclusion for this column ships with Build M (register entry cites this issue); the exclusion lifts with the rebuild, exactly like the miscellaneous_income/#393 pattern.

Start with the staged base builder and its QRF chain for partnership_self_employment_net_earnings, using the export-mass parity gate and the base-j/base-m comparison as context. Compare donor-PUF and imputed positive/negative leg distributions, then use per-target checkpointing to rebuild the affected chain. Done means the signed base mass is no longer near cancellation and the reviewed parity exclusion can be lifted.