Filter non-finite de_sample() draws in sample() for unbounded priors too

[pedroliman]

Fixes #244

Bug

sample.nsbi_posterior() (R/posterior.R) only filters a non-finite draw from the density estimator inside if (bounded) { ... }, where bounded <- !is.null(prior$lower) || !is.null(prior$upper). That guard was added by #234/#236 specifically to work around within_support() returning NA (not FALSE) for a NaN row, and R's matrix indexing keeping rather than dropping a row selected by an NA logical index.

The guard never runs for an unbounded prior (prior_normal(), or any prior_custom()/family without lower/upper) -- the common case; most of the package's own NPE examples use prior_normal(). For an unbounded prior, a NaN/Inf row that an under-trained MAF/NSF/MDN can produce was rbind'd straight into the returned draws matrix with no warning, and attr(draws, "acceptance_rate") still reported 1.0. This silently corrupts summary(), pairplot(), and sbc()/tarp() diagnostics built on sample(), and can surface downstream in map_estimate() as a confusing "theta contains non-finite value" error that blames the seed draw rather than the internal density-estimator artifact that produced it.

Fix

Moved the non-finite-row filter out of if (bounded) so it runs unconditionally, right after de_sample(), using a plain is.finite() row check independent of within_support(). The existing within_support()/rejection-sampling logic stays scoped to bounded exactly as before -- this is additive, not a replacement. Because collected is filtered before acceptance_rate <- nrow(collected) / max(n_tried, 1) is computed, the dropped row is reflected in acceptance_rate for every prior, bounded or not, with no separate accounting needed.

log_prob.nsbi_posterior()'s normalize = TRUE path needs no matching change: its de_sample() call for the acceptance-constant estimate is itself gated on normalize && bounded, so for an unbounded prior that call -- and any chance of it returning a non-finite draw -- never happens in the first place. de_log_prob() is called unconditionally in that function, but only on user-supplied theta, which is already run through check_finite() earlier in the same function; it isn't fed draws from de_sample(), so it needs no guard here.

Tests

Added two regression tests to tests/testthat/test-posterior-nonfinite-de-draw.R (mirroring the existing #234/#236 tests, but with prior_normal() instead of prior_uniform()):

• sample() drops a mocked NaN row from de_sample() and returns the requested n finite draws, running a second sampling batch to make it up.
• When every draw in the one batch is corrupted, sample() warns about the shortfall and acceptance_rate correctly reports 0, not the pre-fix 1.0.

Checks

• R CMD INSTALL --no-docs . && (cd tests && Rscript testthat.R): FAIL 0 | WARN 0 | SKIP 79 | PASS 1373 (torch unavailable in this sandbox, so neural/MDN/MAF/NSF tests skip, as documented in CLAUDE.md; all validation-path tests, including the new ones, ran and passed).
• DESCRIPTION bumped to 0.6.20, NEWS.md entry added (PR number filled in once assigned).
• No exported function's signature or docs changed, so man/*.Rd needs no update.

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.38%. Comparing base (9b7451b) to head (0d540ba).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #245   +/-   ##
=======================================
  Coverage   98.38%   98.38%           
=======================================
  Files          31       31           
  Lines        4034     4035    +1     
=======================================
+ Hits         3969     3970    +1     
  Misses         65       65           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.