feat(core): record per-episode cost in EpisodeReport (#253)

[larstalian]

Self-Review

• PR title uses a conventional commit style summary
• Scope is focused and matches the linked issue or change theme
• I reviewed the diff for architectural drift and unintended public API changes
• Tests and docs were updated where behavior or workflows changed

Closes #253

Summary

EpisodeReport carried the graded result but no notion of what the episode cost to run, so a "capability per unit cost" comparison had to be reconstructed with an external stopwatch.

• Add EpisodeCost(wall_seconds, realize_seconds, turns), populated by the episode loop: perf_counter at realize start and grade end, realize_seconds for the setup (realize + reset) portion, turns counted from record_turn.
• It rides on EpisodeReport as a defaulted field — zero-cost when unset, so the existing direct constructions (SWE, trading, TRL, training tests) are unaffected — and serializes in as_dict for eval runs and the dashboard.
• Backing-agnostic: the same field is populated whether the world ran as PROCESS or any future backing, so cost is comparable across the fidelity range.

Token attribution is deliberately out of scope here — it needs per-backend usage reporting, a separate piece.

Testing

Verified through the real episode loop (scripted solvers, no LLM): turns matches the number of recorded turns (2 for a two-turn solver, 0 for a no-op); wall_seconds >= realize_seconds >= 0; as_dict()["cost"] carries the three keys. The five test files that construct EpisodeReport directly all still pass, confirming the defaulted field is backward-compatible.

Review Notes

• EpisodeCost is exported alongside EpisodeReport at the package top level.
• Wall-clock spans realize → grade (excludes teardown, which is cleanup, not episode work). Token usage is the natural follow-up once a backend reports it.
• Pairs with the backing work (Manifest/RunConfig honor only seed — wire backing + scale selection #251, merged) so cost is measured at every rung of the fidelity range from the first training run.