Prototype U,V/Q-guided multi-peak JAX marginalization

[oshaughnessy-junior]

Scope

Adds an opt-in diagnostic seam for genuinely local joint marginalization over reflected time, two angles, and inverse distance. It does not change production likelihood dispatch.

• ranks distance-following U,V/Q structural starts, reduces sampled symmetry aliases before the cap, then expands each verified group action exactly once
• refines every retained start with bounded sequential JAX Newton/polish steps and a strict stationary gate
• integrates the union of finite full-Hessian affine cells with a full Gaussian-mixture overlap partition
• runs a richer enrichment tier and accepts only when both local diagnostics pass and the log integrals agree within the configured budget
• invokes a lazy finite dense/exact reserve on any decline; planner failure never becomes -inf or point deletion
• retains the failed deterministic global cover only as a frozen optional diagnostic, outside the runtime decision

Adversarial fixes

• symmetry aliases no longer consume representative capacity; the odd targeting-grid phase is preserved because forcing group alignment falsified strict SNR160 convergence
• time endpoints require one-sided local-maximum comparisons
• ordinary affine-cell overlap is partitioned, not blanket-rejected
• a failing reserve is invoked once and is not relabeled as a planner exception
• explicit-array memory telemetry is more conservative and explicitly excludes backend allocator/compiler/AD high-water memory

Real-table gate (1e-3 nat policy)

row	decision	local / oracle error	modes	explicit-array model
22, SNR 40	dense reserve	0 after fallback	4	4.9 MB
22, SNR 160	local accept	-4.3e-6 nat	4	4.9 MB
Lmax=4 HM51	local accept	-7.1e-5 nat	2	19.1 MB

The HM regression preserves the relevant second mode: sampled proxy rank 15 at -37.14 nat, refined contribution at -11.671 nat. No sampled-height pruning is used before refinement.

Validation

• 15 passed in 56.09s in the focused CPU JAX suite, including full real-table 22 low/high-SNR and HM integrals, symmetry-cap, overlap, endpoint, finite-reserve, and frozen-cover regressions
• py_compile, flake8 --ignore=E501,W503, and diff checks pass

GPU device high-water validation remains infrastructure-blocked: the available CUDA/JAX container aborts while creating its client thread under current host thread pressure. The result field therefore labels only modeled explicit arrays, not measured GPU RSS.

[interhost-cross-review-canary-v2]

Independent automated review completed at the recorded exact commit. Detailed findings were withheld from public output by the private-context egress policy and require private human declassification.

[oshaughnessy-junior]

Two notes from testing #273's algebraic phi enumeration against this PR's planner. Neither asks for a change here; the first is a finding in your favour, the second is an availability problem.

1. The U,V/Q lattice is sufficient — a much better start set changes nothing.

At fixed time and distance the exponent x*A - 0.5*x^2*B - 4log x is a 2-D trig polynomial, so the algebraic stationary set applies directly to it (same w = 1 / 2 convention, no adapter). On tables from make_synth -> angle_coefficient_tables at m_max=2 and m_max=4 HM:

case	lattice: starts / worst phi miss	+ algebraic: starts / worst miss	delta lnI	enum cost
m_max=2 rho~18	8 / 0.686 rad	64 / 0.025	+0.000e+00	0.6 s
m_max=2 rho~90	8 / 0.686	64 / 0.026	+9.1e-13	0.6 s
m_max=2 rho~359	8 / 0.686	64 / 0.026	+0.000e+00	0.7 s
m_max=4 rho~18	13 / 0.184	1618 / 0.015	+5.7e-14	135 s
m_max=4 rho~90	13 / 0.090	1601 / 0.005	+0.000e+00	127 s
m_max=4 rho~359	13 / 0.090	1604 / 0.005	+0.000e+00	135 s

18-27x better coverage of the true phi-maxima, identical retained mode count, identical ok, lnI equal to float noise. Your Newton refinement absorbs the difference. #273 is parked as draft on this basis.

One caveat on how far to take that: select_refined_modes retained 1 mode at m_max=4 and 4 at m_max=2 against 10-32 true phi-maxima on these fixtures. If the mode filter is the binding constraint, the test is partly blind to start quality. Is n=1 at m_max=4 the intended behaviour of contribution_cutoff_nats and the dedup tolerances, or is it tighter than you meant?

2. The real-table regressions cannot run outside your host.

_HM_PACKET = /tmp/hm51_Ctables_incl0.6.npz and the two /tmp/rift-paper-av-ladder/... packets are node-local, so test_hm_second_mode_survives_unsafe_proxy_gap, test_hm_two_tier_integral_matches_overcomplete_oracle and the SNR40/160 tests skip everywhere else — including CI. The three headline rows in the PR description are currently unverifiable by anyone but your session. Moving the packets to $HOME (shared across the interactive nodes) or checking in a reduced fixture would make them reproducible.