Add per-domain RewardAdapter #199

Context

A pack's TaskFamily.check_success(...) returns a structured
EpisodeResult — {"success": <bool>, "subgoals": {...}, "reason": <str>}.
Training loops want a scalar (or multi-objective vector) reward.
Today, the only mapping is "did success: true → 1.0, else 0.0", and
that's hardcoded in examples/codex_eval.py's post-processing.

That's fine for a binary cyber task. It's wrong for trading
(continuous P&L), robotics (geometric distance + smoothness +
energy), and most other domains.

What this issue tracks

Add a RewardAdapter ABC and per-pack implementations. Each pack
ships its natural one; training loops consume them.

• Cyber: success → 1.0, plus partial credit from the
  EpisodeResult.subgoals the pentest check already returns (closer
  chains → higher)
• Trading: P&L → Sharpe / drawdown / risk-adjusted return
• Robotics (when it lands): geometric task success + smoothness +
  energy

Why it matters

Without this, every training script has to write its own
result-to-reward mapping, and the mapping leaks domain
knowledge into the trainer. With it, the pack owns the reward
shape and the trainer is domain-agnostic.

Where to start

• openrange-pack-sdk — _types.py defines EpisodeResult
  ({success, subgoals, reason}) and FeasibilityVerdict; _protocols.py
TaskFamily.check_success returns the former. RewardAdapter is
  downstream of EpisodeResult.
• src/openrange/core/episode.py — EpisodeReport.episode_result
  (an EpisodeResult) is what an adapter consumes.
• packs/cyber_webapp/cyber_webapp/families/pentest.py — the cyber
  check_success, for reference on the subgoals structure.
• examples/codex_eval.py — the current ad-hoc mapping; that's
  what an adapter would replace.

Suggested shape

class RewardAdapter(ABC):
    @abstractmethod
    def reward(self, report: EpisodeReport) -> float | Sequence[float]: ...

Pack ships a default (Pack.default_reward_adapter() or similar);
training loops can override.

Acceptance

• RewardAdapter ABC in openrange-pack-sdk (packs import only the
  SDK, not openrange), with a Pack.default_reward_adapter() hook.
• Cyber pack ships a default that returns 1.0/0.0 plus a richer
  signal from EpisodeResult.subgoals (closer chains → higher
  partial credit).
• One runnable example showing the adapter consumed by a training
  script.

Notes

Keep the v0 cyber adapter simple — binary success is fine until a
training run actually needs partial credit. The point of this issue
is the seam, not the perfect reward function.