feat(orchestrator): EnvMixStrategy seam for env selection

[hallerite]

What

Extract TrainSource's weighted round-robin env selection into a swappable EnvMixStrategy seam (default WeightedRoundRobin). Example selection — the per-env reshuffling cursor — stays in TrainSource.

This is slice (b) of the composable algorithm abstraction: cleanly separate which env (global EnvMixStrategy) from which example (per-env, slice c builds on this).

Changes

• orchestrator/sampling.py (new): EnvMixStrategy ABC + WeightedRoundRobin default. pick() returns the next env name via weighted random choice.
• TrainSource delegates the env pick to self.env_mix.pick(); it still owns dataset loading, the per-env cursor, reshuffle-on-exhaustion, and env_costs.

Behavior

Behavior-preserving. WeightedRoundRobin draws from TrainSource's existing RNG (injected), so env selection stays in the same stream as the dataset shuffles — the example sequence is identical to before. (Partitioning RNG per-env is a slice-(c) change, not here.) The public API (TrainSource(train_envs, seed) + next_example) is unchanged.

Testing

• ruff check + ruff format --check clean.
• tests/unit/orchestrator/test_sampling.py (new): WeightedRoundRobin determinism per seed, weight respecting (incl. zero-weight never picked), empty-envs guard.
• tests/unit/test_configs.py (106) pass; imports resolve.

🤖 Generated with Claude Code

---

Note

Low Risk
Refactor with default implementation matching prior logic; TrainSource API and dispatcher wiring unchanged.

Overview
Introduces a swappable env mix seam for training rollouts: global “which env next?” is no longer inlined in TrainSource.

New orchestrator/sampling.py defines EnvMixStrategy and default WeightedRoundRobin, which picks an env name via weighted random choice (same weight rules as before: per-env ratio when all set, else dataset size). TrainSource now calls self.env_mix.pick() instead of rng.choices directly; it still owns datasets, per-env cursors, reshuffle-on-exhaustion, permit costs, and next_example’s public contract.

Behavior-preserving: WeightedRoundRobin uses TrainSource’s existing random.Random instance so env picks stay in the same RNG stream as dataset shuffles.

Tests: tests/unit/orchestrator/test_sampling.py covers determinism, weight behavior (including zero weight), and empty-env validation.

^{Reviewed by Cursor Bugbot for commit 84370ab. Bugbot is set up for automated code reviews on this repo. Configure here.}