feat(benchmarks): add QFUERZA-recovery results for all 5 TS systems

[ericchansen]

Summary

Adds benchmarks/<system>/from-qfuerza/ artifacts for all 5 TS systems from the QFUERZA-recovery validation experiment. Companion to ericchansen/q2mm#290.

What's in each directory

File	Content
validation_results.json	Full run record: provenance (q2mm git SHA, JAX device), starting-point audit (which OPT rows were QFUERZA-overwritten vs published-retained), initial/final objective scores, Seminario R², optimizer trajectory, ratio diagnostics
paper_metrics.json	Published-paper-comparable summary metrics
<system>_optimized.fld	Final optimized force field

Experiment

Each run starts from QFUERZA Hessian-derived bond/angle values (per-molecule Seminario with TS inversion, multi-molecule mean) overlaid on the published OPT topology (frozen partition, vdW, SB, atom-type rows from the published FF). Optimized with SciPy L-BFGS-B + JaxLoss, --ratio-tol -1, run on RTX 5090 (WSL2).

This is not from-scratch FF generation — only 17–33% of active parameters per system are overwritten by QFUERZA. See docs/benchmarks/qfuerza-recovery.md (in the companion PR) for full methodology and interpretation.

Results summary

QFUERZA-start vs published-start final objective:

System	Pub. final OF	QFUERZA final OF	Ratio	Improvement %
rh-enamide	2.70 × 10⁵	2.78 × 10⁵	1.03×	+28.9%
pd-allyl	7.99 × 10⁶	7.98 × 10⁶	1.00×	+0.41%
pd-conjugate	7.24 × 10⁶	8.25 × 10⁶	1.14×	−0.20%
rh-conjugate	5.10 × 10⁶	1.78 × 10⁷	3.49×	+33.4%
heck-relay	1.45 × 10⁶	1.45 × 10⁸	100×	−7.70%

rh-enamide and pd-allyl converge to essentially the same basin as the published-start runs. The other three land in different (worse) basins, with heck-relay failing due to JaxLoss surrogate divergence at the poor starting FF (bond R² = −247 at the QFUERZA starting point).

File size

~530 KB total, all JSON + small .fld files. No large binary artifacts.

Provenance

• q2mm git SHA: 29b61f8 (loader commit on feat/qfuerza-from-scratch)
• q2mm-data git SHA at run time: 72add3d
• JAX device: cuda:0 (RTX 5090, WSL2)
• ratio_tol: -1 (bypassed)