aa: make rank tolerance dim-independent (len·ε·|R₁₁|)

[bodono]

Summary

• The pivoted-QR rank tolerance was aug_rows·ε·|R₁₁| with aug_rows = dim + mem, following LAPACK's max(m,n)·ε·σ₁ convention. For AA this conflates two different dimensions: dim is the caller's state vector length, while the inner LS we're rank-revealing has only len columns (len ≤ mem, typically 5–20).
• Switching to len·ε·|R₁₁| decouples the tolerance from dim. Principled argument: the rank question is about the len columns of [A; √r I_len]; the tall dim rows are the caller's state, not part of the LS's algorithmic dimension.
• Practical consequence: at large state dim the old floor can exceed the noise floor the regularizer itself enforces. With reg=1e-12 and dim=1e6, old tol ≈ 2e-10·|R₁₁| sits well above the √r floor (1e-6·|R₁₁|) that the augmented √r·I rows guarantee by construction — columns the regularizer certifies as healthy could be silently dropped.

Test plan

• Full C suite passes (make LDLIBS="-framework Accelerate" test)
• Full Python suite passes (31/31)
• Bench matches master byte-for-byte on all 14 configs (iter counts, applied/aa_rej/sg_rej, final_err) — at bench's dim range (50–2000) both tolerances are well below |R₁₁|, so no observable change, as expected. This is a large-dim robustness fix, not a perf change.

🤖 Generated with Claude Code