portfolio: make sampling_density the member p_s contract, and give mcsamplerGPU one

[oshaughnessy-junior]

Rebased onto rift_O4d now that #354 has merged. It landed the same three
_pdf_norm[p] = 1.0 hunks independently, so those are dropped here in favour of its
version. What remains is the portfolio contract, reset_sampling, and the tests.

mcsamplerPortfolio builds q_mix from each member's sampling_density().
mcsamplerGPU had none, so any portfolio holding one fell back to the stratified
per-member joint_p_s, which needs every member to report a normalized density.
mcsamplerAdaptiveVolume reports V_s/V, larger by V_s**2, and corrects for it in
its own integrate_log.

Decided: sampling_density() is the contract and must be a normalized density.
joint_p_s stays each sampler's private scale, so AV is unchanged. mcsamplerGPU
gains sampling_density(). reset_sampling now restores _pdf_norm with the pdf,
which #354's fix left stale. The stratified fallback refuses;
portfolio_allow_stratified_density=True restores it.

Production is unaffected

The question here is AV alone, across the normalization changes that just landed.
GW240426 CIP on all.net, 3 replicates at each of three commits:
b281ccff2 (before #354), 94f352ad8 (rift_O4d tip, #354 + #355), and this branch.
Noise floor from same-commit pairs. Every pairwise JS is inside it.

sampler	JS noise floor	pre354 -> tip	tip -> branch	pre354 -> branch
AV (GWTC production), mc	0.0030-0.0071	0.0048	0.0048	0.0042
AV, eta	0.0034-0.0055	0.0045	0.0052	0.0050
default adaptive_cartesian, mc	0.0034-0.0059	0.0058	0.0056	0.0051
adaptive_cartesian_gpu, mc	0.0043-0.0059	0.0054	0.0048	0.0043

AV ln Z: 28.2369 (sd 0.0098), 28.2314 (sd 0.0058), 28.2426 (sd 0.0118). Largest
difference 0.011, one replicate spread.

The defect, and the fix

Constant integrand, non-square box, prior integral 5, exact ln Z = 1.609438,
adaptation running, 16 seeds, same settings both sides:

portfolio	base 94f352ad8	branch
[AV]	1.6094 (sd 0.0000)	1.6094 (sd 0.0000)
[GPU]	1.6134 (sd 0.1158)	1.6134 (sd 0.1158)
[GMM]	1.6207 (sd 0.0093)	1.6207 (sd 0.0093)
[GPU,GPU]	1.5891 (sd 0.0691)	1.5891 (sd 0.0691)
[AV,GMM]	1.6092 (sd 0.0026)	1.6092 (sd 0.0026)
[GPU,GMM]	1.6165 (sd 0.0093)	1.6108 (sd 0.0151)
[AV,GPU]	0.4458 (sd 0.0326)	1.6105 (sd 0.0027)
[GPU,AV]	0.4330 (sd 0.0231)	1.6090 (sd 0.0028)

Five rows are byte-identical. Only the two AV+GPU orderings were broken, by 1.17
nats. [GPU,GMM] is a wash: both members already reported a normalized joint_p_s,
so the stratified fallback was unbiased there. Forming q_mix moves its mean 0.006
toward exact and raises the seed spread from 0.009 to 0.015.

Verified

test/integrators/test_portfolio_member_density.py, 14 cases, in
.travis/test-integrate.sh with the collection-count guard #354 added. Eight
mutations in a separate worktree, one per part of the fix, all caught.

Risk: the refusal raises inside integrate_log, so a member whose sampling_density
returns None late in a run now aborts it instead of finishing with a biased
evidence. No shipped member does that. reset_sampling is called only from tests and
one demo, so that half is a latent fix.

Adjacent, not fixed here

mcsamplerNFlow.draw_simplified returns (rv, p_s, p_prior); every other member
returns (p_s, p_prior, rv). It cannot be a portfolio member today. No nflows here,
so it is unexercised.

🤖 Generated with Claude Code