Stop losing downstream scores to NA, and let splatter/symsim fail properly

[rcannood]

Describe your changes

Stacked on #18 -- please merge that one first. The utils.R changes here sit on top of the compute_logcounts() helper it adds.

Validity fixes for the downstream metrics and two methods. None of these lose or corrupt scores the way #17 and #18 did; they mostly turn results that should have been numbers into numbers.

generate_cosine() was filtering the wrong thing. real[!is.na(real) & !is.na(sim)] operates on the list returned by BivariateMoransI(), not on the Moran's I values inside it -- is.na() on a list of matrices is FALSE for every element, so it never dropped anything. One NaN anywhere then took the whole metric with it: crosscor_cosine came out NA for 32 of 99 runs, while generate_mantel() passes na.rm = TRUE on the same inputs and lost none. Now filters the values:

identical inputs  -> 1        (want 1)
one NaN in sim    -> 1        (was NaN before)
unrelated inputs  -> 0.8618
all NaN           -> NA       (want NA)

A simulation with no spatially variable genes now scores 0, not NA. calculate_precision() returned NA when tp + fp == 0, which is exactly what a simulator producing nothing spatially variable does -- the worst possible outcome, recorded as a missing result. Both negative controls were NA on all 10 datasets, so QC reported 8 of 30 expected control scores and nothing anchored the bottom of the scale.

calculate_recall() deliberately keeps NA in the same situation, and I added a comment saying why: there the empty denominator is a property of the real dataset (brain and hindlimbmuscle have no significant SVGs at all), so there is nothing for any simulator to recover and scoring it 0 would penalise all of them for the dataset.

splatter and symsim now fail properly. Both wrapped their per-cluster loop body in try() and never looked at the result, so a failed cluster left simulated_result a cluster short and the loop carried on; it only fell over later on a length mismatch. And symsim sampled gene lengths with replace = FALSE from the finite gene_len_pool, which errors outright on any dataset with more genes than the pool holds.

logcounts is a double. file_dataset_sp.yaml declared it integer.

Not included

• The precompute components are still dead weight. precompute_downstream writes varm$spatial_variable_genes and varm$spatial_autocorrelation, and sc_features writes uns$L_stats and friends, but the metrics recompute all of it (the reads are commented out in downstream/script.R, with a TODO). I left it alone because wiring it up is not the small change it looks like: the autocorrelation precompute would have to go through the same compute_logcounts() that Normalise the simulated data the same way as the real data #18 introduces or it reintroduces exactly the mismatch that PR fixes, and generate_svg_sparkx() filters mitochondrial genes so its table is not n_vars-aligned. Worth its own PR, or deleting the components. Happy to do either.
• --base tissue is still advertised for methods that don't implement it. Only scdesign3_nb, scdesign3_poisson and srtsim support it; the other five stop(). I had documented that in comp_method.yaml but the edit got reverted underneath me, so I've left the file untouched rather than push it back.
• scdesign3 resource labels. scdesign3_poisson was killed at 8h 3m on breast (exit 140), which is exactly what hightime allows -- but as you said, a method that can't finish in 8h doesn't want a longer wall clock, it wants looking at.

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!

[rcannood]

Same here -- splitting into one PR per fix.