Completes the skill scaffolded in the previous commit (Day 3-5 of the
staged plan).

Added
-----
* scripts/sandbox.py — subprocess sandbox for code candidates:
  - POSIX rlimit caps (CPU, AS, DATA, CORE) applied as best-effort in
    a preexec_fn; each limit failure is swallowed independently so
    macOS's RLIMIT_AS-incompatibility doesn't break the others.
  - Wall-clock timeout via subprocess.run(timeout=...) as a backstop.
  - run_candidate_code() and run_pytest_suite() helpers; the latter
    parses the pytest terse summary into a pass fraction.

* scripts/adapters.py — Tier 2 + Tier 3 wrappers:
  - ExternalEvolverAdapter: lazy shutil.which detection, raises
    AdapterUnavailable with install hint instead of crashing.
  - openevolve_adapter (Apache 2.0, default Tier 2 recommendation).
  - darwinian_evolver_adapter (Imbue, AGPL v3) — subprocess only,
    never imported, so license-viral code never enters the Hermes
    process.
  - export_dspy_jsonl: DSPy-compatible offline records with full
    lineage; default keeps one winner per generation, --all emits every
    candidate.
  - export_gepa_trace: reflective-operator edges (critique_then_edit,
    meta_mutator) in the shape GEPA's trainer expects.

* templates/*.py — five copy-paste-ready fitness templates for prompt,
  regex, SQL, code (uses sandbox), and multi-objective (NSGA-II) runs.

* demos/summarize_10_words/ — end-to-end packaged demo:
  - fitness.py: deterministic scoring (word count proximity, brevity
    keyword, char budget), so the demo runs cheaply without an LLM
    judge and the improvement curve is visible on a small local model.
  - seed/initial.txt, README.md with exact commands and expected
    trajectory.

Changed
-------
* scripts/evolver.py: cmd_export now delegates to adapters.export_*;
  added --all flag for dspy-jsonl exports; imports adapters module.

Tests (tests/skills/test_darwinian_evolver.py)
----------------------------------------------
Expanded from 26 to 39 cases — all green:

* TestSandbox — simple candidate runs, syntax error fails cleanly,
  runaway while-True loop killed within wall-clock+overhead, pytest
  terse-summary parser.
* TestAdapterGracefulAbsence — openevolve + darwinian-evolver adapters
  raise AdapterUnavailable with license-informed install hint when the
  binary is missing (monkeypatched shutil.which).
* TestDspyBridge — default export keeps one record per generation;
  --all mode emits every candidate; GEPA export filters to reflective
  operators and preserves parent/child metadata.
* TestLLMClient — seed propagation verified by intercepting the
  AsyncClient.post body; BudgetLedger records spend across calls and
  raises BudgetExceeded when cap is crossed.
* TestMapElitesCoverage — random 200-sample run fills ≥60 % of a 4×4
  descriptor grid (acceptance checklist NousResearch#4).
* TestEndToEnd — single generation with a fully-mocked LLM produces an
  offspring strictly better than the seed and yields a stable lineage
  hash (acceptance checklist NousResearch#7, plus determinism smoke).

Acceptance checklist status: 10/10 covered.

Repo-level notes
----------------
The original examples/ subdirectory is renamed demos/ because the
repository-level .gitignore lists ``examples/``; keeping the demo as
``demos/`` means it ships to users who install the skill without
requiring an exception in .gitignore.

…dge · critic

Professional-grade upgrade of the skill landed in the previous two
commits on this branch. Four cohesive features that move the skill
from hobby-grade to production:

1. LLM response cache
-----------------------
* New `scripts/cache.py` — content-addressed SQLite cache keyed by a
  blake2b-16 of the normalised request body (model, temp, max_tokens,
  seed, messages). Backed by the existing `lineage.db` so the cache
  ships with the experiment directory.
* `llm.LLMClient` gains an optional `cache` field; `complete()` checks
  the cache before every HTTP call and writes on miss. Cache hits
  short-circuit the network AND the BudgetLedger, so reruns are zero
  cost and the budget accounting stays accurate.
* `evolver run` enables the cache by default; `--no-cache` disables it.
* New subcommand: `evolver cache <dir> stats|purge`.
* Acceptance: `test_end_to_end_replay_is_bit_identical` — a second run
  with the cache populated makes ZERO HTTP calls (monkeypatched
  `httpx.post` raises on invocation) and produces an identical
  `lineage_hash`.

2. Pairwise judge + Bradley-Terry MLE
---------------------------------------
* New `scripts/judge.py` — implements:
  * `aggregate_bradley_terry()` — Hunter 2004 MM iteration (pure
    stdlib, ~30 LOC) with Laplace smoothing for candidates that
    never won or never lost.
  * `sample_pair_schedule()` — round-robin for small populations;
    least-seen-anchor sampling for larger ones. Invariant: every
    candidate appears in at least ceil(rounds × 2 / pop) matches.
  * `PairwiseJudge` — LLM-backed judge with a position-bias guard
    that randomly swaps LEFT / RIGHT per call; decodes first-line
    verdicts robustly (LEFT / RIGHT / TIE with fallbacks).
* `fitness_spec` gains `judge="pairwise"`, `pairwise_rounds=40`.
* `evaluator.evaluate_pairwise()` runs the schedule, records votes
  in `pairwise_votes`, solves the MLE, writes `bt_scores`, and copies
  log-odds onto `Individual.fitness` so the existing selector pipeline
  (tournament / MAP-Elites / (μ+λ)) works unchanged.
* Incompatibility with NSGA-II is enforced explicitly; the runner
  errors out early instead of silently corrupting state.
* Acceptance: `test_condorcet_order_recovered_through_mocked_judge` —
  a judge that prefers lex-higher genomes, combined with the MLE,
  recovers the full six-candidate ranking through 30 pair rounds.

3. Constitutional reward-hacking critic
-----------------------------------------
* New `scripts/critic.py` — `ConstitutionalCritic` runs a second LLM
  over the top-K of every generation with a structured JSON contract
  ({"risk", "evidence", "signal_tags"}).
* `templates/constitution.md` — a default rule-set covering literal
  short-circuits, judge flattery, regex over-matching, test-harness
  exploits, spurious correctness, and brittle templates.
* Penalty applies ONLY to the in-memory `Individual.fitness` that
  feeds selection; the raw `fitness` SQLite row is untouched, keeping
  the audit trail pristine.
* `fitness_spec` gains `critic="on"|"off"`, `critic_threshold`,
  `critic_top_k`, and `critic_model` (cheap-model override).
* Runner hook fires after both seed evaluation and each offspring
  generation; JSON parse failures soft-fail to risk=0 so a noisy
  judge can't crash a run.

4. FastAPI + Plotly dashboard
-------------------------------
* New `scripts/dashboard.py` — read-only FastAPI app with endpoints
  GET /api/{summary, fitness, pareto, lineage/{cid}, operators} and
  WS /api/stream (polls `lineage.db` and pushes a JSON event when
  a new generation lands).
* `templates/dashboard.html` — single-page, vanilla JS, Plotly and
  Mermaid from CDN; no build step, no npm.
* New subcommand: `evolver dashboard <dir> [--host 127.0.0.1 --port 8787]`.
* Binds loopback by default; non-loopback hosts print a warning.
* Graceful absence: `dashboard` command prints an install hint and
  exits non-zero if fastapi/uvicorn aren't installed. Every other
  subcommand continues to work.

Storage schema
--------------
Four new tables, all additive (`CREATE TABLE IF NOT EXISTS`), so v0.1
experiments open and run unchanged:
  * llm_cache (key, response, tokens, model, created_at)
  * pairwise_votes (generation, left, right, winner, seed)
  * bt_scores (candidate, generation, log_odds, iters)
  * critic_evaluations (candidate, generation, risk, evidence,
    signal_tags, model, evaluated_at)

Tests
-----
Four new test files, one per feature — 31 new cases, all green:

  tests/skills/test_darwinian_evolver_v02_cache.py      8 cases
  tests/skills/test_darwinian_evolver_v02_judge.py     12 cases
  tests/skills/test_darwinian_evolver_v02_critic.py     6 cases
  tests/skills/test_darwinian_evolver_v02_dashboard.py  5 cases

Full darwinian-evolver suite: 39 (v0.1) + 31 (v0.2) = 70, all green.
No regression in existing repo tests (full tests/skills/ run: 170/170).

Acceptance checklist (rows 11 + 12 from the plan): both pass.

Invariants maintained
---------------------
* Every v0.1 behaviour is preserved; new features are opt-in (except
  the cache, which is transparent and correctness-preserving).
* `lineage_hash` gains determinism it didn't have before: with the
  cache, identical seeds produce identical hashes across runs.