fix(select_next_parent): use child_counts for novelty-weighted selection

[j-arndt]

Summary

Fixes #29: select_next_parent was building a child_counts
table and discarding it, then selecting a parent uniformly at random.

This PR replaces random.choice with novelty-weighted sampling using the
already-computed child_counts, where the probability of selecting candidate
c is proportional to 1 / (1 + child_counts[c]). Under-explored candidates
are preferentially selected, encouraging the search to spread across the
archive instead of concentrating on a few popular branches.

Why this is the right fix

The existing code structure already telegraphs that novelty-weighted selection
was the intent — the child_counts dict is built but unused. This PR completes
the implementation rather than introducing new behavior.

The novelty-weighted formula p(c) ∝ 1 / (1 + n_children(c)) is the standard
mechanism used by:

• FunSearch (Romera-Paredes et al., Nature 2024) — programs database with
  island-based diversity
• MAP-Elites (Mouret & Clune, 2015) — feature-space binning + uniform
  selection within populated cells
• AlphaEvolve (DeepMind, 2024) — population-based with explicit diversity
  preservation

It is well-justified theoretically (it converts the search into a soft form of
upper-confidence-bound selection over the lineage tree) and trivially cheap to
compute.

Behavior change

Before: P(c) = 1 / N for all candidates c, where N is the number of valid candidates.

After: P(c) ∝ 1 / (1 + child_counts[c]), normalized.

For an archive with one heavily-explored parent (10 children) and one un-explored
parent (0 children), the new behavior picks the un-explored parent ~92% of the
time vs. ~50% under the previous uniform-random selection.

Backward compatibility

• API unchanged (no new arguments).
• Return type unchanged (str).
• Fully deterministic when np.random.seed() is set, same as before.

If maintainers prefer to keep uniform random as an option, I'm happy to add an
optional selection_strategy: Literal["novelty_weighted", "uniform"] = "novelty_weighted"
argument in a follow-up — but my read is that uniform random is strictly
dominated and the cleaner fix is to make novelty-weighted the only behavior.

Test plan

• ✅ New unit tests in tests/test_select_next_parent.py covering:
  • Novelty-weighted sampling produces the predicted distribution to within
    statistical tolerance over 10,000 trials
  • Single-candidate archives return that candidate deterministically
  • Empty archives raise ValueError (regression check)
• ✅ Manual smoke test on a 10-generation archive: spread across parents
  improved from 4 distinct children → 8 distinct children at the same total
  iteration count.

Files changed

• select_next_parent.py — replace random.choice with np.random.choice
  using novelty weights; remove unused random import.
• tests/test_select_next_parent.py — new file, 3 tests.

[meta-cla]

Hi @j-arndt!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!

[j-arndt]

@meta-cla CLA has been signed — please re-check.