hank-limits: analytic-limit and aggregate-identity tests for HANK codes

Tests that a HANK code must pass at limits with known answers, reported as errors and observed convergence orders. Version 0.3.0 adds checks V1-V6 computed from primitive quantities the code reports: the GE target residual of the vanishing-liquidity solve, steady-state market-clearing and budget residuals, income normalisation, the documented specification along equilibrium paths, and the solved exogenous paths against the requested ones, with a verdict that missing or non-finite values cannot hide. Mutation tests of the suite on hand-made and independently generated errors, with pre-registered predictions, are documented in docs/MUTATION_v0_1.md, docs/MUTATION_v0_2.md and docs/CHECKS_v0_3_0_AMENDMENT.md. A numerical audit before the release (docs/NUMERICAL_AUDIT_v0_3_0.md) found no negative mass in the stationary solver and no effect of its absolute value or weight clip on the results; it led to one change, a fixed backward-iteration seed in the permanent-income calibration. Reference implementation on sequence-jacobian 1.0.0, with minimal reproductions of three observations about that library: F1, F2 and F3 (hetinputs and hetoutputs differentiated one-sided, reported as shade-econ/sequence-jacobian#41). Not peer reviewed.

AI use: Developed with extensive assistance from Claude Opus 5.5 (Anthropic) and reviewed with GPT 6.1 Sol (OpenAI); the author directed and reviewed the work and is responsible for the content.