Story: IR curve follow-ups

(Describe the user-visible outcome this story delivers.)

• Matrix math for the multi-factor processes (affine, HJM, LMM, QG) uses Eigen (vcpkg port eigen3, header-only). Prior-art analysis rejected std::mdspan (absent from this toolchain), Boost uBLAS (legacy API, no Cholesky), glm (graphics-only) and NVIDIA matx (CUDA-first). Eigen is the modern CPU dense-algebra standard: value semantics, expression templates, LLT/LLT-with-eigen-fallback Cholesky for correlated shocks.
• Exact one-step Gaussian transitions everywhere one exists (the codebase convention, and QuantLib's own OU evolve): BK on ln r, affine/QG on the factors. Euler is used only where no exact transition exists (LMM's rate dynamics), with its O(dt) bias measured and documented.
• One generic rate_tree lattice in math/ serves every tree model: BDT fills a binomial (2 children, p = 0.5), BK a centered adaptive trinomial (Hull-White moment-matching probabilities, valid for any horizon – no clamping). The processes build the tree; the utility walks it and propagates state prices. BDT extends flat beyond the input curve rather than bounding the horizon (no new limitations at the start).
• QG bond pricing uses the exponential-quadratic recursion (completion of squares over the per-tick shock, Eigen inverse + determinant). At Gamma == 0 it reduces to the affine recursion bit-for-bit – the discrete-time design anchors the two models together.
• affine_term_structure_process covers Gaussian factors only. CIR stays its own process: the square-root factor's exact transition does not fit Gaussian OU dynamics, and forcing it through would break its exact transition. Documented as a future extension.
• Engine work only: none of the six processes is wired into process_factory yet. System-infra wiring is PR 2 (the G2 precedent).
• oresmd::index_family mirrors the SQL CHECK constraint on ores_synthetic_ir_curve_generation_configs_tbl (22 families) as the single source of truth, and is_overnight() encodes its tenor rule (only libor/euribor are term). The seed/config side stays a plain lower-case token; the enum mismatch that broke party provisioning from realistic-2026 seed data is resolved on the oresmd side.
• The dialog's index family combo stays hand-patched (all 22 families, b36839fd0) rather than enum-driven: ores.qt/synthetic already links ores.marketdata.api, but the combo lists user-facing labels that would not gain from magic_enum round-trips; both sides cite the same CHECK.
• lookup_entity is architectural duplication (DQ migration precedent, commits 724253fc8 / d04000791). The drift task migrated the three IAM lookups (tenant_type, account_type, tenant_status) to domain_entity models and deleted the _lookup_entity.org files. The lookup models were the true authority for the tenant-scoped schema; the legacy orgs were the stale duplicates.
• Lookup migrations bind :profile: simple-lookup in * Flags — never hand-authored feature flags. The profile assigns the tenancy, workspace, UUID-PK, change-reason-cache, and pagination flags the party_status blueprint uses, and regenerates with zero diff.
• The tenant-scoped schema is preserved end-to-end by the domain route: tenant_id stays in the PK and gist exclude, with the standard hardening (name unique index, security definer insert function, clock_timestamp(), system-tenant bootstrap check, notify triggers).

• The six engine tasks (Black-Karasinski, BDT, affine, HJM, LMM, QG) moved to the Add new stochastic processes story on 2026-08-11: their engines are merged (PR #1932) but not wired into the system, so the system wiring is tracked there, not here.

Closed at sprint 25 close. Every task is done.