Divide the solver's source filter back out of the recombination

[lispandfound]

Both solvers low-pass their source time functions: SW4 through its
prefilter command, EMOD3D through bfilt with a corner derived from
min_vs and the grid spacing. The wave equation is linear, so filtering
the source is equivalent to filtering every trace -- which means the LF
leg arrives at bb-sim already low-passed, and the matched pair then
low-passes it a second time.

The result is a hole around the merge frequency: the matched high-pass
and low-pass are power-complementary by construction, but only if each
leg is filtered once. tests/test_bb_filters.py measures the hole and
then measures that each correction closes it.

Because the filter is applied to the source rather than the output,
bb-sim can divide it back out here instead of the solver having to be
re-run. --solver reads the filter the realisation actually specifies
(rather than assuming one), and refuses to guess: an SW4 prefilter that
is not a lowpass, or an EMOD3D configuration that also high-passes, is
an error rather than a correction for the wrong thing.

--filter chooses which leg absorbs it:

lf restores the low-frequency leg exactly. Needs the largest boost.
hf fills the missing power from the high-frequency side instead.
Cannot restore below the merge frequency, where the HF leg
carries nothing to scale up.
both one factor on both legs. Best conditioned -- current_power is
bounded below by the high-pass leg -- and restores the power sum
exactly, at the cost of touching both legs.

MAX_BOOST caps any correction at 10x. The two real configurations need
about 2.4x; anything near the cap means the source filter rolls off
faster than the target and the leg is being reconstructed from content
that is not there. warn_if_ill_conditioned says so up front.

Without --solver the recombination is bit-for-bit unchanged: the
correction is applied after the matched pair rather than in place of it,
and the gains are exactly 1. Verified against the previous commit's
output.

What was applied is written into the broadband file's attributes.
Without that the correction is invisible downstream and two files that
differ by it look identical.

Co-Authored-By: Claude Opus 5 noreply@anthropic.com

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}