bayesmith.bridge.jaxns_bridge

The JAXNS nested-sampling adapter: one module, consuming one compiled problem.

R5 Task 6. The backend decision is recorded in the R5 backend
evaluation: jaxns passes
every condition that scores a candidate, blackjax fails condition 1 because it
does not answer ``overflowing_outside_latent``, and §0.6's default drops the
second backend.

**What this module is allowed to see.** A :class:`~bayesmith.compiled.
CompiledEvidenceProblem` and nothing else -- no ``Graph``, asserted by type on
the way in. That class sits in ``bayesmith.compiled`` rather than in
``dispatch/evidence.py`` where it was written, precisely so this import can be
at module scope: ``bridge -> dispatch`` would close a cycle that
``tests/test_layering.py`` raises on. The ruling is in that module.

**What this module does NOT build, and why that departs from plan 6.1.** The
plan says the adapter "returns a populated ``EvidenceResult``". It cannot, and
the reason is a rule this module found by breaking it:
``tests/test_layering.py::test_the_artifact_protocol_is_a_leaf_and_dispatch_is_
what_reaches_it`` asserts that the units reaching ``artifacts`` are exactly
``["dispatch", "evaluation"]``, and says a third "would have to be argued for".
A first draft of this file imported ``EvidenceComponent`` and
``WeightedDrawsPosterior`` at module scope and that assertion failed with
``bridge`` in the list -- which is the test doing precisely what its docstring
promises.

The edge is not argued for, because it is not needed. ``evaluation`` cannot do
its job without artifact types; this module can. It reports NUMBERS, and
``dispatch`` -- which already assembles every other ``EvidenceResult`` in this
package, at ``task.py:3246``, out of facts only the planner has
(``plan_ref``, ``fingerprints``, ``PRODUCER``) -- turns them into artifacts.
That is also what ``bridge/numpyro_bridge.py::nuts`` does: it returns
``dict[str, jax.Array]``, and ``bridge/arviz.py`` imports artifact types inside
a function for the same reason.

**So the split is: jaxns facts here, bayesmith policy in `dispatch`.** The bit
LABELS below are jaxns's own words; the map from a label to a termination
member is this package's judgement, and it is carried here as the enum's plain
string VALUES so that stating it costs no import. The test that it is total --
that every string is a real ``TerminationReason`` and every bit has one -- lives
in ``tests/``, where reaching ``artifacts`` is free.

**The box, and what it costs.** jaxns's ``Prior`` wants a quantile function
``U -> X``, and a compiled problem carries ``log_prior``, ``log_likelihood`` and
``prior_sample`` -- a density is not invertible by inspection and a sampler is
not a transform, so none of the three yields one. The generic route is the
reparametrisation the bake-off measured: declare a bounded box ``B``, give jaxns
a uniform prior over it, and move this package's own prior into the likelihood,

    log L'(x) = log pi(x) + log L(x) + log |B|

so jaxns integrates ``INT_B pi L dx``. Three consequences, each a cost the
verdict already names: ``B`` is an input jaxns needs and the compiled problem
does not carry; the answer is the evidence TRUNCATED to ``B``, so mass outside
it is an error the oracle's own truncation term bounds; and the sampler now
contracts against ``pi L`` rather than ``L``, which is not the problem nested
sampling's efficiency argument is about.

**Two things jaxns does that a caller has to know.** Importing it writes
``jax_enable_x64`` **process-globally** -- reproduced here on 2026-09-07, with
``UserWarning: JAX x64 is not enabled. Setting it now.`` -- against this
package's rule that ``src/`` never touches ``jax.config``. §0.16's owner
decision attaches to this module because the backend it was conditional on was
chosen. And ``TerminationCondition()`` **disables every stopping rule**: each
field defaults to ``None``, ``None`` means "off", and the real defaults live in
``jaxns/public.py``, which builds one only when it is handed none. This module
never constructs one and never passes ``term_cond`` at all.