feat(identification): combined zero-and-sign restrictions (ARW construction)

[thomaspinder]

Summary

New ZeroSignRestriction identification scheme: named impact-zero constraints and horizon-specific sign constraints in one scheme, via the Arias-Rubio-Ramírez-Waggoner recursive null-space construction — not a penalty approximation.

• Construction: per draw, columns are built recursively — the orthonormal null basis of the stacked zero rows plus previously-drawn columns, hit with a spherical Gaussian and normalised — giving exact zeros (SVD precision) and exact orthogonality by construction; sign restrictions are then accept/reject with the whole candidate abandoned on any failure (resuming mid-candidate would be a silent distribution bug; commented at the site and guarded by the equivalence test).
• Regularity: the Rubio-Ramírez-Waggoner-Zha counting condition (z_j ≤ n − j on the sorted padded sequence) is checked deterministically at identify() entry — impossible specifications fail with a named error before any sampling, distinct from low-acceptance runs, which warn with counts and diagnostics (zero_sign_acceptance_rate, failed fractions, mean attempts, max zero violation).
• Failure policy: unaccepted draws are NaN'd (never silently fall back to the Cholesky factor — that would violate the scheme's zero guarantee; a deliberate, documented divergence from SignRestriction), with on_failure="raise" opt-in.
• Validation anchors: full triangular zeros reproduce |chol(Σ)| bit-exactly (acceptance 1.0), and with positive-diagonal signs, chol(Σ) itself; an n=2 single-zero case matches the closed-form column analytically; signs-only mode is distributionally equivalent to SignRestriction (acceptance diff 0.0000).
• Honest scoping, stated in four places (docstring, explanation page, how-to, CONTEXT.md): draws are the standard recursive construction without ARW's volume-element importance weights (required for uniform-conditional set-identified inference; filed as a follow-up rather than shipping an unverifiable formula), per-θ retry follows the in-house convention, and the sampler is Haar on O(n).
• Docs: explanation section with the counting rule, a climate how-to (an activity shock with zero contemporaneous impact on temperature — physical-lag contemporaneous exogeneity), reference entry, CONTEXT.md, ARW 2018 + RWZ 2010 bib entries.

SignRestriction is byte-untouched; the class is appended at end-of-file to stay out of the #183/#200/#208 conflict zones.

Closes #144

Review

Planned by a Fable-tier planning agent; independently reviewed (verdict: approve, nothing blocking) — the reviewer re-derived the recursion, re-implemented it from scratch with scipy's null_space and reproduced all inspected cell distributions within MC error, verified the early-abandon path has exactly one failure exit, and confirmed the bit-exactness argument for the Cholesky anchor.

Tests

35 tests: the degenerate-Cholesky anchor, the n=2 closed form, generic zero/Σ/sign satisfaction to machine precision, permuted-ordering invisibility, RWZ rejection, the failure policy three ways, diagnostics contract, and the SignRestriction equivalence. Fast suite: 562 passed, 29 deselected; ruff/ty clean.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Egjd7ToFeb9TQqFnfRQZxV

[codecov-commenter]

Codecov Report

❌ Patch coverage is 96.66667% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 95.1%. Comparing base (866ca8f) to head (55956bc).

Files with missing lines	Patch %	Lines
src/impulso/identification.py	96.6%	3 Missing and 3 partials ⚠️

Additional details and impacted files

@@          Coverage Diff           @@
##            main    #218    +/-   ##
======================================
  Coverage   95.0%   95.1%            
======================================
  Files         45      45            
  Lines       3099    3279   +180     
  Branches     380     417    +37     
======================================
+ Hits        2945    3119   +174     
- Misses       111     114     +3     
- Partials      43      46     +3     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.