[flyingdoubleG]

Problem

ProgramDatabase._add() stores the candidate in self.programs before running the novelty check (the check looks the program up by id), but when the check fails it returns program.id without undoing that store:

self.programs[program.id] = program
feature_coords = self._calculate_feature_coords(program)
if not self._is_novel(program.id, island_idx):
    return program.id  # Do not add non-novel program

So a program the judge rejected as NOT_NOVEL is still in the database. It is not on any island, but it

• can be picked by _sample_random_parent() (and therefore by sample()),
• shows up in global get_top_programs() / get_best_program(metric=...),
• counts towards population_size — _enforce_population_limit() is skipped on this path, so the database can grow past the limit,
• is written to checkpoints by save(),
• has already widened the min/max feature scaling via _calculate_feature_coords().

Since add() returns the id, callers can't tell it was rejected either: ProcessParallelController stores its artifacts, logs it to the evolution trace as an accepted child and increments the generation counter of current_island (the rejected program has no island metadata) instead of the target island.

This path only became reachable after #483 fixed the NOT_NOVEL parsing; before that the judge never rejected anything.

Minimal repro (population_size=3, one island, identical embeddings, judge always answers NOT_NOVEL):

add returned dup0 ... dup4
islands:  [{'seed'}]
programs: ['dup0', 'dup1', 'dup2', 'dup3', 'dup4', 'seed']

Fix

• On novelty rejection, remove the provisional entry (restoring any previous entry with the same id) and return None. This is the contract add() already has for population_strategy.admit rejections, and both existing callers handle it (ProcessParallelController raises _CandidateRejected, _migrate_one checks is not None).
• Compute feature coordinates after the novelty check so rejected programs don't change feature_stats.
• Docstring / log message updated to mention novelty rejection.

Tests

New tests/test_novelty_rejection.py (mocked embedding client + a judge that always returns NOT_NOVEL): rejected programs are not stored, never sampled or ranked, not checkpointed, and don't touch feature_stats. All 4 fail on main and pass with this change. Existing novelty / migration / population-strategy tests still pass; the rest of the unit suite has the same results as main locally.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}