US frame clips PUF Schedule C losses to zero (puf_tax_detail nonnegative_outputs) #1238

The puf_tax_detail stage lists self_employment_income_before_lsr under nonnegative_outputs (packages/microcosm-build/src/microcosm/build/us/spec/sources.yaml, around line 328). It arrived in f85c248 ("F0 D3: generate complete US spec bundle") with no stated rationale. The donor's self_employment_income is the signed E00900 (policyengine-us-data datasets/puf/puf.py: puf["self_employment_income"] = puf.E00900), and farm operations income on the same stage is already re-derived "without clipping".

Effect on the certified Route A release (populace-us-2024-0fb05b6-4b57d15a287c-20260930T150401Z):

Why it matters. HT2 and district-file N00900 count every return with a Schedule C entry, profit or loss. Since #1228 / #1230 binds them as such, the frame can meet that count only with profit carriers. The #1230 re-solve (experiments/1228-net-income-less-loss/) measures the result: 28.1M profit units against IRS 22.0M, with profits at 0.87×. A positive-part binding gets profits right but overstates the net 1.3×. Only a frame with losses gets both.

What a fix has to check:

• every consumer that assumes nonnegative Schedule C income: SE tax, the QBI/SSTB split in us_runtime/qbi_inputs.py, the self-employed health insurance attribution, earnings bases in puf_support.py, EITC earned income, and is_self_employed;
• the weighted loss count and amount against Table 1.4, on a rebuilt stage.

Evidence: ~/PolicyEngine/_reviews/soi-net-income-less-loss-20261012/ (scripts/carriers.py, out/carriers_*.json).