Weight-threshold prune does not remove zero-weight samples with checks enabled

[Jammy2211]

Overview

The sample weight-threshold prune does not appear to remove zero-weight samples from samples.csv, even with validation checks enabled. Found while reproducing #1486; split out because the blast radius is different — this affects every saved samples.csv, not one tutorial.

Evidence

Two local runs of autogalaxy_workspace/scripts/guides/results/_quick_fit.py (capped Nautilus, n_live=100, n_like_max=300):

run	env	rows in samples.csv	zero-weight rows	prune logged?
A	smoke profile (PYAUTO_SKIP_CHECKS=1)	300	299	no
B	JAX on, PYAUTO_SKIP_CHECKS unset	300	299	no

config/output.yaml sets samples_weight_threshold: 1.0e-10 in both the workspace and the packaged PyAutoFit config. samples_above_weight_threshold_from (samples.py:481-523) keeps only sample.weight > weight_threshold, so 299 rows with weight == 0.0 should have been dropped in run B.

The weight column is unambiguous — collections.Counter over it gives exactly {'0.0': 299, '1.0': 1}.

Run A is explained: samples.py:505-506 nulls the threshold under skip_checks(), which is a known and intended behaviour of the smoke profile. Run B is not explained.

The "Samples with weight less than ... removed from samples.csv." log line (samples.py:517-520) never fired in either run.

Suspect (unconfirmed)

updater.py:212-215 returns early, before the prune at line 220, whenever the summary instance raises:

try:
    instance = samples_summary.instance
except exc.FitException:
    return samples, samples_summary, None, samples   # <-- skips save_samples + prune

In this fixture the stored samples include points PyAutoGalaxy's validate_ell_comps rejects (that is #1486), so samples_summary.instance plausibly raises FitException and takes this branch — meaning an invalid sample anywhere in the list silently disables pruning for the whole run. That would make the two issues mutually reinforcing.

Not verified. Alternatives not ruled out: the final save going through a different path, or log_message being suppressed by _disable_output while the prune does run.

Plan

• Confirm whether updater.py:212-215 is the path taken, e.g. by instrumenting or by constructing a fixture whose summary instance is valid while other samples are not.
• If confirmed, decide whether an unreconstructable summary instance should really forfeit the prune and the samples_summary write, or whether the early return should be narrowed.
• Add a regression test asserting that zero-weight samples are absent from the saved samples when skip_checks() is false.

Context worth carrying

This fixture reports f_live=1.0000, N_eff=1 — the n_like_max=300 cap means the sampler never converges, so the weight vector is degenerate (one-hot) by construction. Do not reason about "typical" weight distributions from it.

Split out of #1486 (fix(autofit): reconstructing a stored sample raises through ignore_assertions=True), which has the full reproduction.