DOOP corpus: 21 native-validated applications

Semantic correction: These historical results used an unguarded Object[] store shortcut. CPU/GPU agreement here does not establish upstream DOOP equivalence. The corrected unary guard and Eclipse proof are at https://github.com/harp-lab/srdatalog-python/releases/tag/doop-object-array-guard-v1 . Other datasets have not been revalidated under the corrected query. Historical timings/profiles are not corrected-query performance results. Original evidence assets remain unchanged.

Native admission of 21 distinct DOOP application workloads: 6 small / 5 medium / 6 large / 4 xlarge. The canonical query and normalization semantics are unchanged.

All21 applications pass native SRDatalog bitmap execution on RTX 6000 Ada with device-only cuda_async allocation: normal exit, all74 counts and all37 complete CPU IDB tuple sets per application (777 exact comparisons). Zero warmups/one repeat validate correctness, not speedups.

Replaces H2O with Groovy2.4.21 and Kotlin1.5.31 with genuine Kotlin1.4.32 (719,597,196 VPT rows). The v1 release remains immutable; other selected inputs retain their pinned existing archive URLs. H2O and old Kotlin retirement evidence distinguishes unavoidable live capacity from fixable copies and scratch fragmentation.

Baseline audit:20 native exact passes, combining19 earlier ownership-fixed pool runs and current Kotlin cuda_async. Jython baseline still exceeds VRAM while sorting3,026,761,641 transient NEW rows at iteration113; its bitmap plan passes. No hidden fallback or reduced facts.

Assets include raw replacement archives, extraction provenance and preparation proofs, complete CPU/native reports, baseline audit, exact Jython allocation trace, selection/retirement evidence, compiler verification and SHA256SUMS. Compiler fixes are in PR#94, runtime commit d1fe748. No facts or binaries are committed to Git.