arch-calibrated RC crashes: microcalibrate non-finite loss on negative/tiny-support Arch targets

[MaxGhenis]

arch-calibrated RC crashes: microcalibrate non-finite loss on negative / tiny-support Arch targets

The arch_calibrated recalibration (microcalibrate backend, Arch NIPA+SOI targets from #200, from the post-imputation checkpoint) crashes:

ValueError: Relative error contains non-finite values (NaN or Inf).
Check the inputs: targets near -1, zero initial weights, or Inf/NaN from the estimate function.
  microcalibrate/utils/metrics.py:44 in loss

(artifacts ..._arch_calibrated_ce1e1b9_20260604/logs/recalibrate_microcalibrate_arch_from_post_imputation...log)

Cause

Many Arch targets are large negative values at ~0.01% support — the log warns on dozens, e.g.:

• arch_target_5075 = -6.74e8, arch_target_4913 = -1.57e9, arch_target_5111 = -2.67e9, arch_target_4697 = -3.12e9 — each "supported by only 0.01% of records", initial estimate ~-1e5 (3–5 orders of magnitude off).

These are signed income-loss components (net capital losses, partnership/farm/rental losses — #209 added signed rental/farm income). microcalibrate's loss is (estimate - target)/target, which is non-finite for negative / near-zero / tiny-support targets (the "targets near -1" hint, after normalization).

Fix (any one unblocks the arch-calibrated RC → promotion)

1. Exclude tiny-support targets (e.g. support < ~0.5%) and/or negative-loss targets from the microcalibrate set — they're statistically unstable at 0.01% support regardless (microcalibrate already warns). Likely a --min-support / target-filter in the arch-recalibrate CLI.
2. Make microcalibrate's loss sign-robust — normalize by max(|target|, floor) or use abs/Huber on the relative error instead of dividing by a signed/near-zero target.
3. Route signed income-loss aggregates through entropy (handles them; the prior RCs did) or keep them audit-only, calibrating only positive/magnitude aggregates.

Status

ce1e1b9 crashed; 8fe13c9 fell back to backend=entropy. No build running. Last good artifacts: the collapse-entropy RC + merged dividend #202 + prior-year #204. Detail in mp-ecps loop journal iter86.