feat(v0.2.2): Pareto selection, novelty filter, backend ensemble & cascade evaluation

[weijt606]

Summary

Brings four search-strategy techniques from recent open-source evolutionary-agent
projects into PolyHarness, plus observability to make them visible. Every feature is
opt-in, dependency-free, deterministic/reproducible, and keeps existing behavior
unchanged — in line with the project's principles (open-source, no security risk,
reasonable & reproducible design).

What's new

1. Pareto-frontier parent selection — parent_selection: pareto

Samples parents from the set of per-task winners instead of always branching from
the single overall-best candidate, keeping specialists alive as stepping stones to
avoid premature convergence. Reuses the per-task scores already in the search log —
no new data collected. (GEPA, arXiv:2507.19457)

2. Code novelty rejection — novelty_filter / novelty_threshold / novelty_max_retries

Detects near-duplicate candidates via stdlib difflib and skips their evaluation to
save API/compute budget. Off by default. (ShinkaEvolve, arXiv:2509.19349)

3. Adaptive backend ensemble — proposer.ensemble / ph run --ensemble a,b,c

A UCB1 bandit picks a backend per iteration and shifts picks toward backends that
produce improving candidates. Fully deterministic (no RNG); prints a per-backend
picks/improve-rate summary. Leverages the existing 8-backend support.
(ShinkaEvolve adaptive LLM-ensemble selection)

4. Cascade evaluation — evaluator.cascade / cascade_threshold / cascade_stage1

Scores a cheap first subset of tasks and only runs the rest if it clears the gate,
saving budget on weak candidates. Per-task mode only; the base harness is always
scored in full. Off by default. (AlphaEvolve/OpenEvolve cascade)

Plus

• Reproducibility — search.seed makes tournament/pareto/novelty repeatable.
• Observability — ph log marks Pareto-frontier members (◆); ph leaderboard
  adds Pareto + Backend columns (Backend only when an ensemble was used).
  SearchLog.pareto_win_counts() is the shared source of truth.
• proposer_backend recorded in each candidate's metadata.json.
• Removed 3 byte-identical duplicate files that tripped ruff N999.

Design notes

• Backward compatible: all flags default to prior behavior; an injected single
  proposer disables the bandit. No public API breakage.
• No new dependencies: novelty (difflib), bandit (math), cascade — all stdlib.
• No new attack surface: the bandit only selects among already-configured
  backends; the sandbox/proposer boundaries are unchanged.

Testing

• ruff check src/ tests/ — clean
• pytest tests/ — 206 passed (173 → 206; +33)
• End-to-end smoke on the bundled math-word-problems template with the offline
  local backend for pareto + novelty + ensemble + cascade.