portfolio: fix --save-samples 0-row export (cumsum linear weights + -inf guard)

[oshaughnessy-junior]

Problem

mcsamplerPortfolio.py --save-samples exported 0 sample rows for peaked (low-to-moderate n_eff) portfolio runs — even at n_eff≈9.

Step 2 of the _rvs cleanup paired the sorted index with the log-weight ln_wt[k] and cumsummed that, so the "cumulative weight" was never a cumulative probability. Any -inf ln_wt entry (rejected/underflowed sample) poisoned the cumsum (cum_sum[-1] = -inf → nan on normalize), and the cum_sum > deltaP mask kept 0 rows. The block was inherited from the original ccd70ff7 draft (copied from the ensemble pattern) but dropped the linear-weight step.

Fix

• Pair/cumsum the linear weight wt[k] (already computed one line earlier), matching the canonical mcsampler/mcsamplerEnsemble truncation.
• Map any non-finite log-weight to -inf (→ zero linear weight), with a degenerate fallback that keeps all rows if no finite-weight sample survives.
• Confined to the portfolio Step 2 block — AV/GMM standalone samplers are untouched.

Verified numerically: reproduces canonical linear-weight index selection exactly on finite peaked input, and keeps all real samples (dropping only zero-weight rejects) when -inf entries are present (old code kept 0).

Also (docs only, no XML schema change)

The sparse sim_inspiral XML from --save-samples carries lnL only and cannot be reweighted by likelihood for a weighted-posterior/shape check. Added a code comment + updated --save-samples help text pointing users to the ASCII --extrinsic-proposal-output (full log-weight) / --calibration-export-posterior.

Validation note

The shape-recovery merge gate (S250114ax best-fit point, --save-samples --save-P 0) needs GPU + data and was not run in the authoring environment — please run it before merge.

🤖 Generated with Claude Code

[oshaughnessy-junior]

Auto-review: not ready to merge while CI is red.

The three failed jobs (docs, install (3.10), install (3.11)) have empty step metadata and their GitHub logs are unavailable (404), so I cannot attribute the failure to this change; the simultaneous pre-step failures look runner/service-side rather than a test assertion. Please rerun CI and require a green result.

Local review: clean merge with rift_O4d; changed Python compiles; the focused test_portfolio_oracle.py completes under the project’s optional-plugin-isolation setup. The direct unisolated run remains blocked by the pre-existing eager NF entry-point import when torch is absent (already tracked separately). The required base-vs-candidate integrator shape gate is running; I will update this review with that result.

[oshaughnessy-junior]

Integrator-gate update: this PR cannot yet satisfy the mandatory portfolio shape gate. The candidate run reports ModuleNotFoundError: No module named 'torch' for every portfolio row, due to the eager optional-NF entry-point import. That is the unresolved optional-plugin-isolation issue tracked in #33, but it makes a base/candidate comparison non-probative for this portfolio-changing patch.

Please merge or depend on the #33 isolation fix, then rerun the standard CPU base-vs-candidate gate (with portfolio strict) and rerun CI. Until then, the red CI plus invalid portfolio gate block merge.

[oshaughnessy-junior]

Closing as stale/already upstream.