DOOP: add CPU/GPU benchmark matrix and exact oracle comparison

[StarGazerM]

Scope

Run the twelve-dataset corpus against the actual canonical Python query, with separately selected CPU, GPU baseline and opt-in GPU bitmap execution.

• Translate the instantiated logical program conservatively to Soufflé, preserving recursion, multi-head rules, negation, supported filters and execution-only SPLIT semantics.
• Build/run GPU workers in fresh processes; require synchronized unlimited fixedpoints, checked lifecycle calls and normal process exit. Kill only owned child process groups on timeout.
• Add checked native cardinality/TSV export APIs using authoritative canonical index columns, inverse permutations and bounded host buffers; handle two-level FULL/HEAD storage without a second full GPU reconstruction.
• Record separate build/load/run/export timing scopes and complete failure artifacts. Defaults: one warmup, three measured repetitions. No silent dataset skipping.
• Record all 74 relation counts and export all 37 derived tuple sets. Distinguish 39 prepared inputs from the 37 inputs actually consumed by the canonical program.
• Require matching query/input identities, complete coverage and exact sorted integer tuple sets against a reference; counts alone are insufficient. Without a reference, correctness is explicitly not_compared.
• Reject empty-root benchmarks. Preserve completed execution evidence when reference comparison itself fails.

Verification

• Integrated Python suite: 1,110 passed, 5 skipped; Ruff checks/formatting passed. A subsequent native-artifact isolation fix passed all three worker-process regressions (including a new regression reproduced failing before the fix).
• All 12 real pinned datasets prepared successfully; each selected exactly one main method. Complete input inventory was measured, not inferred from archive sizes.
• All twelve CPU datasets completed normally, each with all 74 relation counts and all 37 IDB exports. The canonical query produced 2,742,413 VPT tuples for Xalan through 438,477,770 for Jython; H2O consumed 33,024,599 input rows. These are measured canonical-query counts, not assertions of equality with the upstream guide. Validation used zero warmups/one repetition amid other validation activity and is not a performance comparison.
• Complete baseline and bitmap DOOP GPU libraries compiled with clang++-20/CUDA 12.9.
• Bounded native CUDA smoke passed exact tuple export checks for 10 relations / 37 tuples, including EDB export, column permutations and both recursive bitmap source roles.
• Additional native boundary smoke: baseline and bitmap each matched 77,874 expected tuples exactly, exercising multiword bitmaps, negative int32 values, overlapping projections and the 65,536-row export chunk; normal process exit with a 256 MiB pool cap.
• Reproduced the reference-comparison exception accounting bug before the fix and verified completed evidence survives afterward; empty-root rejection was also exercised.
• A real complete Xalan GPU attempt in each mode exceeded a 1 GiB RMM pool cap before fixedpoint. Both recorded failure, with abnormal teardown exit; neither is counted as a successful benchmark or tuple match.
• That failure exposed an allocator log written into the checkout. Workers now run from their external output directory; the artifact-isolation regression fails before the fix and passes afterward.

Validation blocker / draft status

An unrelated vLLM workload occupies most of the RTX 6000 Ada's VRAM (44,742 MiB used / 49,140 MiB total when checked). It was not stopped. Full twelve-dataset GPU equivalence and performance are NOT established; leave this PR draft until that matrix can run with sufficient GPU headroom. No published speedup claim.

Reproduce

See docs/benchmarks.md for preparation, CPU reference, GPU baseline/bitmap comparison, selection, compiler prerequisites, timing boundaries and process-timeout semantics. Keep data/build/results outside the checkout and use a new output directory per suite run.

Stack

Depends on #92, which follows #91. Review and merge in that order. No research archive, generated binary, fact file or result dump is included.