perf: direct-ndtr fast path for TruncatedGaussianPrior.value_for

[Jammy2211]

Summary

TruncatedGaussianPrior.value_for is the #1 cProfile hotspot identified
by the graphical-ep-scale-up
baseline: 33% of total wall time in a graphical Dynesty fit at N=10 and
~16% of EP at N=10 are spent inside scipy.stats._distn_infrastructure.cdf
— invoked from norm.cdf / norm.ppf inside this method. The actual
erf math is cheap; the cost is the scipy.stats Python-side wrapper.

This PR replaces both the prior's and TruncatedNormalMessage's
value_for bodies with a call to a new shared helper
(autofit.mapper.prior._erf_helpers.truncated_normal_value_for) that
uses scipy.special.ndtr / ndtri (and jax.scipy.special.ndtr /
ndtri on the JAX branch) directly. These are the same Cephes routines
that scipy.stats.norm.cdf / ppf wrap — bit-exact equivalent, just
without the dispatch.

Measured on the autofit_workspace_developer toy 1D Gaussian baseline:

Workload	Pre-fix	Post-fix	Speedup
graphical N=3	22.8 s	5.6 s	4.04× (75% reduction)
EP N=3	251.9 s	76.3 s	3.30× (70% reduction)

The measured speedups are 2–4× larger than the cProfile-projected 30%
(graphical) / 17% (EP) because cProfile itself adds wrapper overhead
that masked the true scipy.stats cost in the baseline measurement.

Sanity blocks PASS on both fits; max log L matches the pre-fix baseline
within Dynesty stochastic noise (~1e-3 relative — same scale as
re-running with a different random seed).

API Changes

None — internal implementation change only. TruncatedGaussianPrior.value_for
and TruncatedNormalMessage.value_for retain identical signatures and
docstring behaviour. Their bodies now delegate to the new private helper
autofit.mapper.prior._erf_helpers.truncated_normal_value_for, which is
not part of the public API.

See full details below.

Test Plan

• pytest test_autofit/mapper/prior/ test_autofit/messages/ — 261 passed (includes new bit-exact equivalence test against the OLD scipy.stats.norm.cdf/ppf composition on a (mean, sigma, lower, upper) × unit grid).
• autofit_workspace_developer/graphical/fit.py --total_datasets=3 — sanity PASS, 4.04× speedup vs committed baseline.
• autofit_workspace_developer/ep/fit.py --total_datasets=3 — sanity PASS, 3.30× speedup.
• JAX parity test — JAX path matches NumPy path within 1e-9 relative.
• TruncatedNormalMessage.value_for bit-exact to TruncatedGaussianPrior.value_for for matching parameters (both share helper).

Full API Changes (for automation & release notes)

Added

• autofit.mapper.prior._erf_helpers.truncated_normal_value_for(unit, mean, sigma, lower_limit, upper_limit, xp=np) — internal helper, not exported via autofit.__init__. Computes the truncated-normal inverse CDF via scipy.special.ndtr / ndtri.

Changed Behaviour

• autofit.TruncatedGaussianPrior.value_for(unit, xp=np) — implementation only; signature and outputs unchanged. The function now calls the helper directly instead of going through scipy.stats.norm.cdf / ppf (numpy) or jax.scipy.stats.norm.cdf / ppf (JAX).
• autofit.messages.truncated_normal.TruncatedNormalMessage.value_for(unit, xp=np) — same change as above.

Removed

• None.

Renamed

• None.

Changed Signature

• None.

Migration

• None required. The change is binary-compatible: callers see identical numerical results (the new path uses scipy.special.ndtr / ndtri, which are the underlying Cephes routines that scipy.stats.norm.cdf / ppf already wrap).

🤖 Generated with Claude Code