test: publish 0xE9 cross-model microbenchmark results

[zhoubot]

Summary

Publish reproducible 0xE9 cross-model microbenchmark artifacts and independent ASL corpus inputs.

• uniform 8192-byte cross_model_result ABI plus absolute size symbol;
• verified main and cross_model_stop symbols for direct boot;
• 372 rebuilt ELF artifacts, all e_machine=0xE9;
• host-derived golden and pto-asl-elf-sidecar-v1 generators for every active family;
• integer vector TDIV/TREM divisors normalized away from zero before regeneration.

Corpus

All 341 active microbenchmark/coverage.json cases now have verified corpus rows:

• scalar: 124
• memory: 14
• cube: 11
• vector: 129
• fixp: 63

Every row validates ELF/golden SHA-256, exact LLVM/LLD/TileOp/PTO/source identity, writable PT_LOAD result placement, cross_model_result size 8192, absolute cross_model_result_size=8192, and the stop symbol. Every golden is exactly 8192 bytes.

Reproducibility

• LLVM/LLD: 27c2ff9dce92629ccf9b66140d171a13ce2c0153
• TileOp: 155f610c838e918e1fc30ae44f631a85c440b76b
• PTO-SPEC: e51cf7382584a367b0222411966576971c74d264
• ELF machine: 0xE9

Runtime status

The real scalar add_i32_lat ELF passes the ASL model with result 3905 and an exact 8192-byte golden. The parameterized GTest suite is running with ctest -j8; remaining family results will be added before this draft is marked ready.

Tracks #99.

[zhoubot]

Added eacd2ba to align the runnable corpus with PTO v0.58.5 and the corrected ASL runner path:

• active corpus is now 333 cases; the 8 TPART* binaries are retained as explicit unsupported coverage because those operations are deleted/reserved in PTO v0.58.5
• CUBE and FIXP sidecars now allow up to 1,000,000 architectural instruction steps, avoiding false PTO_STEP_LIMIT failures
• fresh compile gates: vector 121/121, CUBE 11/11, FIXP 94/94
• the two pure expand FP32 cases pass the ASL functional model after the TileOP destination-geometry correction

Dependencies: LLVM PR #78 and TileOP API PR #44. This PR remains unmerged pending the refreshed 333-case ASL run.

[zhoubot]

Owner review requested; this PR intentionally remains Draft and must not be merged yet.

Current signed candidate: c8577e91ed1c66099bef3a70a4b17dbd8e7608b2. Independent code review found no remaining producer/verifier defects after the latest fixes:

• preserves current main benchmark and RES_CHECK behavior; corpus publication is explicit and isolated
• locks the corpus at 333 active + 29 explicitly unsupported cases
• validates ELF64/LE/e_machine=0xE9, result ABI, writable segments, sidecar paths/hashes, producer/toolchain/PTO/asl-model/TileOp identities
• rejects missing, duplicate, mixed-identity, malformed, or hash-mismatched inputs fail-closed
• commits no generated ELF/sidecar/golden/manifest artifacts

Validation completed: 21 producer/verifier/benchmark-stability tests, generation idempotence, Python compilation, Ruff, and git diff --check.

Not yet release evidence: the real 333-case compiler-to-ASL execution has not run. Please review the code and artifact contract; keep the PR Draft until exact toolchain/model execution produces complete immutable evidence.