Fix ks_statistic_gene_cell metrics being dropped from the results

[rcannood]

Describe your changes

Rebuilt on top of 3301ded. ks::kde.test() stays the metric, as @ycao6928 intended -- this PR makes the plumbing around it correct.

In the last full run (27Klprxf2344VO, run_2026-07-11_18-00-02) every ks_statistic_gene_cell metric came out as 100% missing -- 0 results where 110 were expected, for all 28 of them.

The cutoff. extract_uns_metadata only keeps a uns entry with at most uns_length_cutoff elements, and run_benchmark passed 15. The component emits 28 metric_ids, so the ids were thrown away and the report had no way to attribute the values. Raised to 50.

$tstat is not a field. ks::kde.test() returns Tstat, zstat, pvalue, ..., so result$tstat was NULL and c() dropped it silently:

success path -> names: Tstat, zstat, pvalue
  result$zstat = 2.113228  | result$tstat is NULL: TRUE
  c(result$zstat, result$tstat) has length 1 instead of 2

On the success path only 14 of 28 values came out; on the fallback path all 28 -- so metric_values had a length anywhere between 14 and 28 against 28 metric_ids. The cutoff was hiding this, and raising it on its own would have zipped the two together positionally and mislabelled scores rather than losing them. That's why both changes are in one PR.

Failure was the best possible outcome. zstat and Tstat are unbounded -- zstat spans about -6 to +113 across the last run -- while the empirical KS fallback lives in 0..1 and the penalty was a literal 1. With maximize: false both rank above nearly every genuine result, so a component that crashed scored better than one that worked. There's no finite value that reads as "worst" on an unbounded scale, so failures now report NA and show up as missing in the QC instead.

Which brings back the NAs you were fixing, so I fixed those on the kde scale instead. Restored the deterministic jitter retry that ks_statistic_sc_features still uses -- kde.test trips over ties, which count data has plenty of, and nudging the inputs through 1e-8 .. 1e-4 gets it through. A tied input that used to error now returns zstat 4.07. Plus a guard for inputs with no variance: both sides get the same deterministic noise, so a constant input stays identical after jittering and kde.test then calls it far more alike than chance -- a top score for degenerate data. Those report NA.

Behaviour after this PR, checked against the component's own image:

normal 1-D             n=2  zstat=1.024
heavy ties             n=2  zstat=4.073      <- used to be an error
normal 2-D             n=2  zstat=1.027
constant both          n=2  zstat=NA
constant one           n=2  zstat=NA
2-D with constant col  n=2  zstat=NA

and on realistic data, an identical (positive-control) pair gives zstat = -0.8 against +1.2 for a mildly wrong simulator, consistent with the -1.246 we saw for the positive control in the real run.

Config left alone: 28 metrics, min: -Inf / max: +Inf / maximize: false all still match what the component emits.

One thing worth being aware of rather than fixing here: zstat is a standardised test statistic, so at a fixed effect size it grows with n -- mean 2.70 at n=200 up to 25.76 at n=5000 over 20 seeds. Across 10 datasets of different sizes it partly ranks by dataset size. Happy to open a separate issue if you want to think about that.

Checklist before requesting a review

• I have performed a self-review of my code

• Check the correct box. Does this PR contain:

  • Breaking changes
  • New functionality
  • Major changes
  • Minor changes
  • Bug fixes

• Proposed changes are described in the CHANGELOG.md

• CI Tests succeed and look good!

[ycao6928]

Thanks @rcannood !