Design doc: EpisodeResult ↔ trajectory format seam

[larstalian]

The design-doc-first half of #198, split out so it's independently
claimable. Specify how an OpenRange episode maps to an
open-trajectory-gym trajectory. No implementation — just the seam shape.

What OpenRange emits

From src/openrange/core/episode.py:

• per step: Observation (visible_state, events) + the harness-supplied
  AgentTurn (message, tool_calls, tool_results)
• terminal: EpisodeReport.episode_result: EpisodeResult —
  {success, subgoals, reason}, structured and never a scalar
  (openrange-pack-sdk _types.py)

To specify

1. Observation encoding — what an observation is when the surface is
   HTTP routes / MCP tools / files (map visible_state + events).
2. Action encoding — AgentTurn.tool_calls → trajectory action.
3. Reward — defer to the RewardAdapter from Add per-domain RewardAdapter #199 (EpisodeResult →
   scalar / vector); do not bake a scalar into the trajectory.
4. done / terminal — from the episode's terminal_reason.
5. Runtime shape — fresh-world-per-rollout vs. pooled; curriculum via
   auto_evolve(...).

Acceptance

An ADR or DESIGN.md section answering (1)–(5), reviewed and merged.
Unblocks #198 (implementation); pairs with #199 (reward) and #200
(curriculum loop). Read open-trajectory-gym's trajectory format first.