perf(evolution): atomic-worker architecture for parallel proposals (full GEPA parity)

[KE7]

Summary

Replaces helix's three-stage parallel pipeline (Step 1b parent-eval pool → Step 2 LLM pool → Step 3 sequential child-eval) with a single ThreadPoolExecutor whose workers each execute the full GEPA execute_proposal shape atomically — achieving full structural parity with GEPA reflective_mutation.py:268,308,369,420.

Design

Each atomic worker runs:

1. parent_eval — _cached_evaluate_batch(parent, subsample_ids, None, ...) (bypasses cache, mirrors GEPA RM:268)
2. skip-perfect check — if all parent scores ≥ threshold → return SkippedResult (mirrors GEPA RM:308-327)
3. LLM mutation — mutate(parent, eval_for_mutate, new_id, ...) (mirrors GEPA RM:369)
4. tamper check — _detect_evaluator_tamper(child, manifest, config, project_root) (thread-safe read-only)
5. child_eval — _cached_evaluate_batch(child, subsample_ids, minibatch_cache, ...) (mirrors GEPA RM:420)

Budget charges and frontier updates remain sequential in the acceptance loop (apply_proposal_output parity, GEPA RM:472).

Source: /Users/ke/helix-gepa-parity-investigation.md §7 D1.

Result type

@dataclass
class _ProposalResult:
    kind: Literal["success", "skipped", "tampered", "llm_failed"]
    presample_ctx: tuple[...]
    parent_eval_result: EvalResult | None
    child: Candidate | None = None
    child_eval_result: EvalResult | None = None
    tampered_paths: list[str] | None = None
    parent_n_uncached: int = 0
    child_n_uncached: int = 0
    child_usage: UsageStats | None = None

Diff stat

File	Insertions	Deletions
src/helix/evolution.py	458	444
tests/unit/test_evolution_minibatch.py	354	0

Test results

• 866 passed, 0 failed (860 original + 6 new TestAtomicProposalWorker)
• New test class covers: atomicity (same thread), skip-perfect isolation, worker exception isolation, n=3 parallelism, sequential budget charges, tamper-check rejection

E2E validation

Backend	Exit	Cost	Mutations	Accepted	Wall-clock	Errors
Claude	0	$0.11	1 (g1-s1)	1/1 (1.0)	~10s	none
OpenCode	0	$0.00	2 (g1-s1, g1-s2)	2/2 (1.0)	~14s	none

Both runs used num_parallel_proposals=2. The OpenCode run demonstrates the parallelism directly: both worker threads printed "⟳ Running train evaluation…" simultaneously, and both mutations were accepted (both fixed the add_one off-by-one bug independently).

Design invariants preserved

• ✅ Acceptance order = pre-sampling order (not worker completion order)
• ✅ Budget charges sequential (acceptance loop only)
• ✅ Frontier/lineage writes sequential (acceptance loop only)
• ✅ PromptArtifactCollisionError re-raised from worker (fatal, not swallowed)
• ✅ max_workers cap applied to single pool (min(n, config.evolution.max_workers))
• ✅ No force-push, no --no-verify

🤖 Generated with Claude Code