integrators: portfolio GMM dead-member fix (n_comp=None) + AV draw-density repair (bin-ordered head slice)

[oshaughnessy-junior]

Portfolio GMM dead-member fix + AV draw-density repair

Two coupled silent-degradation bugs found by the shape-recovery merge gate (test/expensive_before_merging/integrators/, merged in #31), fixed together because the first was masking the second.

Bug 1: portfolio GMM members never trained (dead member)

mcsamplerPortfolio.setup() forwards kwargs that lack n_comp; mcsamplerEnsemble.setup() defaulted n_comp=None; update_sampling_prior silently no-ops for n_comp=None. Net effect in default production wiring (--sampler-method portfolio without explicit GMM args): the GMM member never trained and every portfolio ran as AV-only — no error, no warning. Discovered by the PR #28 freeze-territory probe (the freeze machinery was governing a corpse).

Fix: setup() defaults n_comp=None → 1 with a loud notice (n_comp=0 remains the documented off-switch); update_sampling_prior warns once instead of silently skipping; plus an up-shape guard for bare (2,) bounds rows in uncorrelated groups (same guard _sample()/q-scoring already had) — without it, fixing Bug 1 alone turns the silent no-op into a crash in default (per-dimension) configuration.

Bug 2: AV draw_simplified head-sliced a bin-ordered cloud (density lie)

sample_from_bins emits points grouped in lexicographic bin order; draw_simplified returned rv[:n_to_get], i.e. only the first ~50–60% of the live-volume bins, while sampling_density (hence the portfolio balance-heuristic q_mix) claimed uniform coverage of all occupied bins. In any multi-member portfolio this systematically biased the recovered shape (pulls up to 0.64σ at d=2, dose-response with take-fraction; it also explains the +0.1-nat residual portfolio evidence bias seen in the #26 validation and the universal negative-pull signature on base). The dead GMM member of Bug 1 partially masked it — a live member re-weights exactly the region AV never populates.

Fix: random subsample (without replacement) instead of the head slice. Validation probe: pull 0.644 → 0.0013, lnZ bias −0.299 → −0.006, JS 0.081 → 1e-5, n_eff unchanged. AV standalone paths are untouched (draw_simplified serves portfolio members).

Validation

• Regression test test/integrators/test_portfolio_gmm_member_trains.py: default-wired AV+GMM portfolio must train its GMM member; n_comp=0 must still disable. Both pass.
• Differential shape gate (standard preset, base rift_O4d @ 841bcfe vs this branch, portfolio held strict): portfolio d2 5/6 IMPROVED(fail→pass) — evidence bias +0.02..+0.12 → |bias| ≤ 0.007 nats, pulls 0.03–0.16σ → ≤ 0.007σ; three more targets starved→pass at d4/d6; AV and GMM standalone rows unchanged. One flagged row (mix_d2_n1_s303) is a threshold-straddler: n_eff 3022→1502 trips the n_eff<0.5×base heuristic at 0.497 while every shape metric improves (bias +0.085→−0.002, pull 0.045→0.001) — waived.
• Full artifacts: RIFT_roboto_paper analyses/integrator_shape_gate/gmm_member_fix_gate/ (gate JSONs + comparisons + the complete diagnosis/ probe ladder isolating Bug 2: jitter check, single-member portfolio controls, coverage instrumentation, monkey-patch validation).

Note for reviewers: the gate's two GMM d6_n3 verdict flips between rounds were shown to be OpenBLAS-pthreads jitter on n_eff≈100-marginal starved rows (bitwise-reproducible on an idle node), not code effects.

🤖 Generated with Claude Code

[oshaughnessy-junior]

Auto-review finding: the new regression test cannot collect in a torch-free installation, because importing mcsamplerPortfolio eagerly loads the optional NF entry point and mcsamplerNFlow raises ModuleNotFoundError: No module named 'torch'.\n\nReproduction from this PR checkout:\n\nGW_SURROGATE='' RIFT_LOWLATENCY=1 PYTHONPATH=MonteCarloMarginalizeCode/Code pytest -q MonteCarloMarginalizeCode/Code/test/integrators/test_portfolio_gmm_member_trains.py\n\nThis is not a failure of the GMM regression itself: with entry-point discovery isolated, both the default-GMM-training and n_comp=0 checks pass. But #33 targets rift_O4d directly and does not contain the optional-plugin isolation from #28, so it cannot make this test reliably runnable in the production-style no-torch environment. Please either include that isolation here, make #33 explicitly depend on a merged equivalent, or arrange the test to avoid eager optional-NF discovery.

[oshaughnessy-junior]

Fixed in 8ca6368: guarded each pipeline.load() in the portfolio plugin discovery (skip-with-notice on unavailable optional deps, matching #28's isolation) and converted the regression script to proper pytest functions. Verified with the reported invocation plus a torch-import blocker on PYTHONPATH: 2 passed. The guard makes #33 independent of #28's merge order.

🤖 Generated with Claude Code