feat: data-informed SV priors + dynamics-aware log-vol level (closes #65 #66)

[thomaspinder]

Stacked on #69 — review/merge #69 first; this PR's base is fix/constant-cache-cholesky-factor.

Summary

Closes #65 and #66 in one bundled PR (they touch the same per-variable loop in src/impulso/sv/spec.py).

#65 — multivariate SV priors. The adapter was calling build_priors(np.zeros(T)), which produced a near-pinpoint prior on mu_i (sigma ≈ 1e-6) and a wildly off-target h0_mu ≈ -18.4. Fix:

• VolatilityProcess.build_pymc_latent gains optional data: np.ndarray | None = None.
• VAR._build_pymc_model pre-fits OLS once (cheap relative to MCMC) and passes the demeaned residuals (T_eff, n_vars) as data=resid.
• Multivariate SV adapter loops per-variable, calling build_priors(data[:, i]) so each variable's priors come from its own innovation series.
• Constant accepts and ignores data (Protocol parity).
• Multivariate SV raises ValueError if data is None or wrong shape — the silent zeros-fallback that caused Multivariate SV prior defaults to np.zeros(T) — needs data-informed priors #65 is gone.

#66 — redundant log-vol intercept. The multivariate adapter added an outer mu_i shift on top of the dynamics' own intercept, double-counting the level for AR(1). Fix:

• SVDynamics Protocol gains has_explicit_level: bool (ClassVar — true class-level constant).
• AR1.has_explicit_level = True (its alpha carries the log-vol level); RandomWalk.has_explicit_level = False (no intrinsic level).
• Multivariate SV adapter skips the outer mu_i when dynamics.has_explicit_level.

Standalone univariate SV is unaffected. It goes through _build_pymc_model, not build_pymc_latent. The byte-for-byte regression file (tests/test_sv_standalone_unchanged.py) was removed pre-existing in commit 034abe0, but the standalone smoke tests (test_sv_spec.py::test_sv_spec_fit_random_walk_smoke / _fit_ar1_smoke) pass unchanged.

Test plan

• pytest -m "not slow" -q — 236 passed, 22 deselected
• pytest -m slow tests/test_volatility.py::TestPosteriorEquivalence -v — 2 passed (Constant byte-for-byte unchanged)
• pytest -m slow tests/test_sv_spec.py::test_sv_spec_fit_random_walk_smoke tests/test_sv_spec.py::test_sv_spec_fit_ar1_smoke -v — 2 passed (standalone SV unchanged)
• pytest -m slow tests/test_fitted.py::TestFittedVarSigmaDispatch::test_sv_sigma_returns_per_t -v — 1 passed (SV-VAR end-to-end)
• ruff check src/ tests/ — clean
• ruff format --check src/ tests/ — clean
• ty check — clean
• New tests added:
  • TestHasExplicitLevel (2) — SVDynamics adapter trait
  • TestConstantBuildPymcLatent::test_accepts_and_ignores_data_kwarg — Constant Protocol parity
  • TestSVBuildPymcLatentDataValidation (2) — multivariate SV data=None and wrong-shape guards
  • TestSVMultivariatePerVariablePriors::test_calls_build_priors_with_per_variable_slice — spy on build_priors per variable
  • TestSVMultivariateDynamicsAwareLevel (2) — AR(1) skips v{i}_mu, RW keeps it
  • TestVarPassesResidualsToVolatility — CapturingVolatility integration test pinning the threading contract

Commits

1. ffb13b7 — has_explicit_level on SVDynamics
2. a0b8964 — data kwarg on Protocol + Constant
3. 2873b9a — multivariate SV requires data; VAR threads OLS residuals
4. 2359289 — per-variable priors + dynamics-aware level (core fix)
5. 1593218 — CapturingVolatility integration test

Notes / follow-ups (out of scope)

• P3's SV-VAR recovery test (TestSVVarRecovery) is currently softened to shape + finiteness only. With this fix, h-recovery should improve materially; a follow-up commit could re-strengthen the assertions. Deliberately deferred so this PR stays scoped.

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.4%. Comparing base (1d9fd87) to head (d91eeef).
⚠️ Report is 1 commits behind head on fix/constant-cache-cholesky-factor.

Additional details and impacted files

@@                         Coverage Diff                          @@
##           fix/constant-cache-cholesky-factor     #70     +/-   ##
====================================================================
+ Coverage                                92.3%   92.4%   +0.1%     
====================================================================
  Files                                      26      26             
  Lines                                    1126    1141     +15     
  Branches                                  107     111      +4     
====================================================================
+ Hits                                     1040    1055     +15     
  Misses                                     65      65             
  Partials                                   21      21             

☔ View full report in Codecov by Sentry.
📢 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.