[shinaoka]

Closes #161.

Fermions.jl fails on v2 with UndefVarError: default_sampling_points not defined in SparseIR. It calls SparseIR.default_sampling_points(basis.sve_result.u, length(basis)), a v1 internal. In v2 it cannot be restored: an SVEResult holds only the singular values (libsparseir does not return the SVE functions), so basis.sve_result.u fails as well. The fix therefore goes into the caller, and this PR documents it.

Changes

• README.md: a "Migrating from SparseIR.jl v1" section. A table maps each v1 call that no longer runs to its v2 replacement (required ε, SVEResult, default_sampling_points, uhat_full, roots/sign_changes/find_extrema, factorize, .τ, TauSampling64, custom kernels), followed by the two behavior changes that raise no error (the order of given Matsubara points is kept; basis.u(τ) accepts τ < 0).
• llms.txt (new): a short v2 API overview for LLM agents, plus the same migration list.
• test/spir/migration_tests.jl: checks that the documented replacement of default_sampling_points(sve.u, L) / (sve.v, L) gives L ascending points in (-1, 1) at which U_L / V_L of a larger basis vanish.

Verification

• Compared against SparseIR v1.1.4: 2 .* SparseIR.default_tau_sampling_points(basis) ./ β .- 1 matches v1's default_sampling_points(basis.sve_result.u, length(basis)) to ≤ 1e-9 (the SVE accuracy for ε = 1e-8) for F/B, Λ = 1, 40, 1000. The basis[1:L] variants for τ, ω and Matsubara points (incl. positive_only) match v1 for L = 5, 7, 12, 25.
• Every v1 call in the table was run against v2 and fails as described.
• Ran the Fermions.jl propagator tests with local SparseIR.jl v2 and the one-line change in src/Propagators/model.jl: prop_model (7), lehmann (18), gauss (35), prop (964), prop_schedule (10) pass.

🤖 Generated with Claude Code