Balance DOOP corpus across 21 measured applications

[StarGazerM]

Change

Expand the catalog from 12 to 21 distinct applications, with 6 small / 5 medium / 6 large / 4 xlarge instead of 5/2/4/1. The size thresholds are unchanged. Sixteen genuinely new application families were extracted and measured; nine were selected rather than padding the suite with more small cases.

Tier	New applications and canonical VarPointsTo rows
small	Bloat 11,227,250
medium	Clojure 16,591,860; Chart 17,221,612; javac 24,163,325
large	JRuby 60,512,570; PDFBox 69,315,812
xlarge	Soot 412,802,921; Scala 612,741,889; Kotlin 1,076,040,872

• Keep all original twelve archive URLs, byte sizes, checksums and tiers unchanged.
• Use canonical CPU VPT cardinality uniformly across all 21 entries; retain old upstream counts separately for traceability.
• Move source provenance to individual entries instead of assigning every application the Chopin/HuggingFace identity.
• Preserve all raw extracted facts and genuine application roots. No duplicated application versions, synthetic roots, multiplied facts, sampling or threshold changes.
• Leave the canonical query, normalizer and CPU/GPU implementations unchanged.

Published data and evidence

Corpus prerelease: nine pinned raw archives, per-application source/dependency/image/platform hashes and extraction commands, extraction logs, full CPU count reports/input fingerprints, candidate-selection audit, preparation-equivalence evidence and SHA256SUMS. Data and binaries stay outside Git. Total catalog downloads are approximately 2.96 GB; complete raw inputs approximately 52.3 GB.

Verification

• Every selected application completed the unchanged canonical CPU fixedpoint: 74 relation counts and all 37 IDB exports. Original twelve results are reused only with the identical canonical query hash and their immutable validated inputs; all nine additions were newly executed.
• Re-prepared all nine new ZIP archives through the real CLI into a fresh root: every one of the 39 normalized relation records, metadata hash, entrypoint set and query hash matches the CPU-validated snapshot. Original CPU report/manifest hashes were not rewritten.
• After publishing, downloaded all nine new archives through the normal public fetch command into a fresh root, without the archive-cache option. All size/SHA256 checks passed.
• Exercised list and list --tier xlarge; the latter selects Soot, Jython, Scala and Kotlin.
• PYTHONPATH=src python -m pytest -q: 1111 passed, 5 skipped, one existing Jupyter deprecation warning.
• ruff check examples/doop_benchmark.py: passed.

Limits

Zero warmups/one repeat establish cardinality and completion, not comparative performance. GPU equivalence/performance is not claimed here. Static extraction retains its reflection/phantom limitations: Kotlin has 23 phantom methods and 3 phantom-based methods after completing declared dependencies plus JNA and Java 8 tools; Scala also retains phantom-based methods. Exact diagnostics are preserved, not suppressed. Failed extractor runs, including runs that exited zero with Soot errors, were not promoted. Upstream software retains its own licenses.