Autoresearch cumulative mutations on functions.cpp (train sr=0.4491)

[thibautbar]

Single-file changes to FIRE7/sources/functions.cpp synthesized by a Claude-driven autoresearch loop on a held-out feynman-ibp-bench train split (56,102 records, 13 topologies). All 100% valid (modular at prime 2017). Test eval at this state: sr=0.310 @ 100% validity (911/913 records covered, 2 SIGABRT crashes on 15D outliers — to investigate).

Diff: +575 / -29 lines, all in one file. Concentrated in:

• forward_stage() (per-sector Laporta driver) — most of the change
• improve_ibps() (IBP reordering)
• calculate_product() (small helper)

The 16 kept iters compose two cumulative ideas:

I. EXPANDED SYMMETRY SEEDING (sr ramp 1.000 -> 0.613)
iter 1 finer-grained early termination in level loop
iter 2 pure-dot symmetry seeding at all dot levels
iter 3 + extend to one-numerator levels
iter 4 + extend to two-numerator levels
iter 5 drop the numerator cap entirely
iter 6 pre-IBP early-exit (probe done BEFORE the IBP pass)
iter 8 front-load write_symmetries() for ALL pending levels up front
iter 10 recover "sending-higher" internal-symmetry relations
iter 11 + extend front-loading to upper-neighbour levels
iter 12 seed-diagonal upper-neighbour level

II. CLOSURE-AWARE EARLY EXIT (sr ramp 0.613 -> 0.449)
iter 24 within-level early-exit (file-scope hook)
iter 25 within-GROUP early-exit (finer than #24)
iter 29 within-group refinement
iter 33 depth-2 symmetry composition (revisit resolved points)
iter 37 cross-group closure segregation (stable_partition for closure-first)
iter 39 multi-wave closure re-segregation (deepen #37)

The driving constraint: every kept mutation must preserve macro_validity == 1.0 on the full train split. Several agent attempts (iters 28, 40) hit lower sr but broke validity (~1% wrong reductions) and were correctly discarded.

Each commit on claude/main_exp in the bench repo contains the agent's rationale and a comparison.md showing per-topology before/after.