Preserve rare-tail carriers and persist realized regimes in receipts

[MaxGhenis]

Rare signed tails: preserve deleted carriers and persist realized regimes (no-regression infrastructure)

From the battery adjudication's rare-signed-tails package: all 48 red QED checks classified by recomputed realized regime (35 zero-inflated-positive, 13 three-sign; 17 early / 31 late transfer ownership), with the key mechanism finding that Keogh's two native carriers were deleted before transfer fitting — it is not structurally absent, so the declared-absence route is rejected and the carrier union is preserved under the unchanged cap and seed. Realized regimes are now persisted through transfer banks, receipts, H5 checkpoints, and resume validation (closing the ACS-transfer receipt gap the adjudication named).

Honest evidence framing: the cold 1% host build shows 127 = 127 failure lines vs baseline — exactly zero behavioral movement, because at 1% the retirement donor (~1,080 rows) is below the 5,000-row cap and the preservation path is structurally inert (committed: experiments/battery_rare_signed_tails/verify_1pct.md). What merges here is verified no-regression infrastructure plus the receipts contract; the behavioral demonstration binds at the 25% rung, and the remaining 40 shape checks need held-out calibration evidence per the committed classification (experiments/battery_rare_signed_tails/REPORT.md, realized_regimes.json).

Suite green (5,973 build + all shards); no tuning; no threshold changes.

Authored by sol; verified and graduated by Claude.

🤖 Generated with Claude Code