Benchmark patched MPO contraction crossover

[shinaoka]

Summary

• add case 6 for global versus partitioned chain-TreeTN MPO-MPO fit contraction
• cache deterministic global MPO inputs in HDF5 outside timed regions
• pin merged tensor4all-rs revision 9e9aeda
• record the cap-128 crossover sweep on AMD EPYC 7713P
• make the top README table link directly to every benchmark's details, runner, report, and plots/results

Result

At N=2048, one core, two runs, patch cap 128:

input χ	global [s]	patched [s]	global / patched
192	11.791	13.353	0.88×
224	15.468	13.103	1.18×
256	19.606	12.274	1.60×

The crossover is between χ 192 and 224. χ 128/192/224 are explicitly documented as input-cap-limited points. Cap 64 is excluded from future case-6 measurements.

Full report: result/linux-epyc-7713p/mpo_mpo_aniso_patched.md.

Accuracy status

All recorded case-6 points exceed the unchanged 1e-4 sanity gate under the ITensor-style cutoff. The runner exits nonzero after writing each record; the report states this explicitly. No tolerance or gate was relaxed.

Verification

• cargo test --release — 39 tests across all targets
• cargo clippy --all-targets -- -D warnings
• cargo check --locked --all-targets
• uv run scripts/report.py result/linux-epyc-7713p
• README local-link validation
• cargo fmt --all -- --check
• cross-model final review: reviewer-flash-opencode-go — Correct-to-merge