Expose AA trust_factor in ScsSettings (opt-in, default off)

[bodono]

Summary

Adds acceleration_trust_factor to ScsSettings. When set to a finite positive value, the AA solve activates two coupled mechanisms (introduced in cvxgrp/aa#54):

1. Trust region: each AA solve rejects steps where ‖D γ‖_2 > trust_factor · ‖g‖_2.
2. Adaptive regularization: the Tikhonov term r adapts via accept / reject feedback (start large = AA ≈ DRS, shrink on accept, grow on rejection).

Together these recover the type-II broken regime on slow-contraction maps. The default is INFINITY (= "no cap") — existing callers see no behavior change.

Why opt-in?

Setting acceleration_trust_factor = 10 on the Maros-Meszaros + Netlib benchmark:

Config	Default (HEAD)	trust_factor=10	Δ
Type-II penalized geomean	32.8	14.5	−56% ✓
Type-II solved	164/231	195/231	+31 ✓
Type-I penalized geomean	13.7	16.0	+17% ❌
Type-I solved	196/231	193/231	−3

The trust region helps type-II dramatically but slightly hurts type-I, and the historical best (type-I at 13.7) is still the global best — so flipping the default would be a small regression for users on type-I. Leaving INFINITY as the default preserves current performance for everyone while letting type-II users (and anyone willing to sweep) opt in.

What's in this PR

• include/scs.h: acceleration_trust_factor field on ScsSettings.
• include/glbopts.h: AA_TRUST_FACTOR default = INFINITY.
• src/util.c: default set in scs_set_default_settings.
• src/scs.c: validate() rejects NaN / non-positive (INFINITY accepted); aa_init call now passes trust_factor.
• src/rw.c: serialized at the END of the settings block; reader tolerates EOF there (data files written by older scs versions load fine, with the field defaulting to INFINITY).
• src/aa.c, include/aa.h: vendored from cvxgrp/aa#54 — adds trust_factor parameter to aa_init plus the trust-region + adaptive-r mechanisms.
• test/problems/test_solver_options.h, test/run_tests.c: sweep test (INFINITY + 10.0) and validation test (NaN / 0 / negative rejected).

Test plan

• make test: 59/59 pass on both direct and indirect builds (was 57; +2 new tests)
• No-op when acceleration_trust_factor is left at default (INFINITY) — random_prob and other read-from-file tests still load and solve correctly
• Benchmark validated against bodono/solver_benchmarks Maros-Meszaros + Netlib on 231 problems: type-II recovery numbers as quoted above

Dependency

Depends on cvxgrp/aa#54 landing first (the vendored aa.c / aa.h in this PR matches that branch).

🤖 Generated with Claude Code