_runtime_amp_failsafe's SUSPECT label is uninformative under vmap (i.e. on every production path)

[oshaughnessy-junior]

_runtime_amp_failsafe's SUSPECT label is uninformative under vmap — i.e. on every production path

Summary

anglemarg._runtime_amp_failsafe records a "this call's amplitude exceeded the
size the dense grids were built for" label via jax.debug.callback, guarded by
jax.lax.cond. Under vmap a batched predicate lowers to a select, so both
branches execute, and the callback branch passes a literal True rather
than the predicate — so _record_amp_failsafe cannot distinguish tripped from
not-tripped.

Every --mode flowmc-* run with --angle-marg-scheme exact or laplace is
therefore labelled SUSPECT regardless of whether it is, and a genuinely
undersized run is indistinguishable from a sound one. The label is
uninformative in both directions on exactly the path it exists to guard.

Measured

Sound run, predicate false by four orders of magnitude (amp_sizing forced to
1e12 against a worst reported amplitude of 10.49):

transformation	tripped	n_calls
plain call	False	0
under jit	False	0
under vmap	True	2

Reproduction: build any likelihood data, call _runtime_amp_failsafe three ways
(plain / jit / vmap) with a deliberately huge amp_sizing, and read
anglemarg._AMP_FAILSAFE after jax.effects_barrier().

Why it matters beyond the label

Two independent sessions were on the point of using the label's cleanliness as
evidence of a method's soundness for paper 1. It cannot serve that purpose. A
label named SUSPECT that is True on a sound run is worse than no label,
because it invites exactly that use.

Suggested fix

Evaluate the intended predicate on the HOST — compare the max reported amplitude
against AMP_FAILSAFE_TRIP_FACTOR * amp_sizing there — or pass the traced
predicate into the callback instead of a literal, so the record reflects it.

Also, adjacent and probably the same fix

The comment block at anglemarg.py ~:570-577 promises "a POISON term the caller
ADDS to its result ... FAIL CLOSED". The function returns None, both call
sites discard it, and the block immediately below retracts the design and
explains why. Two adjacent contradictory blocks, the first of which reads as a
contract.

Provenance

Pre-existing; NOT introduced by #221. Found independently by the chip-05
session and confirmed by the measurement above during #221's review.
Deliberately not folded into #221 (pre-existing, needs its own validation).
DESIGN_jax_distance_quadrature.md section 5(a) records it and instructs
readers to treat the fail-safe as ABSENT on any vmapped path — but a design-note
caveat does not stop the next person building a soundness check on the label,
which is why this is filed.