Replace RL trajectory extraction with episode views

[cm2435]

Summary

• Replace legacy RL trajectory extraction with sample-scoped RL episode/read views that derive task/spawn lineage from the task tree and worker bindings.
• Add actor-state, episode, and projection services for RL consumers, with minimal DTOs shared across the view layer.
• Change rollout polling to emit training records instead of legacy trajectories, and update the TRL adapter/example to demonstrate single-agent TRL where subagent spawns are treated as environment stochasticity.
• Regenerate REST OpenAPI/contracts for the new rollout response shape.

Stack-shaped review map

PR1: RL view contracts and actor state

• Adds ergon_core.core.views.rl.models plus actor-state read support.
• Tests: ergon_core/tests/unit/views/rl/test_actor_state.py.

PR2: Episode tree reconstruction

• Adds RlEpisodeReadService for sample-level episode trees, attempts, steps, task parentage, worker binding, actor identity, and experiment/environment lineage.
• Tests: ergon_core/tests/unit/views/rl/test_episode_view.py and ergon_core/tests/integration/rl/test_rl_episode_view_roundtrip.py.

PR3: RL training projections

• Adds RlProjectionService that normalizes sample-level reward and emits per-attempt training records from the episode view.
• Tests: ergon_core/tests/unit/views/rl/test_projections.py.

PR4: Rollout API integration

• Deletes core/rl/extraction.py and routes rollout polling through the episode/projection services.
• Updates generated REST contracts to expose trainingRecords.
• Tests: ergon_core/tests/unit/rl/test_rollout_training_records.py, ergon_core/tests/unit/rest_api/test_rollouts_di.py, and architecture guard updates.

PR5: TRL example adapter

• Updates ergon_infra.adapters.trl_http for training-record payloads and moves the runnable TRL script into examples/training.
• Leaves ergon_infra.training.trl_runner as a deprecation shim only.
• Tests: ergon_infra/tests/unit/test_training_experiment_submission.py and tests/examples/test_trl_single_agent_stochastic_imports.py.

Local verification

• uv run pytest ergon_core/tests/unit/views/rl ergon_core/tests/integration/rl ergon_core/tests/unit/rl ergon_core/tests/unit/rest_api/test_rollouts_di.py ergon_core/tests/unit/architecture/test_no_legacy_compatibility.py ergon_infra/tests/unit/test_training_experiment_submission.py tests/examples/test_trl_single_agent_stochastic_imports.py -q -> 40 passed
• uv run ergon test full unit -- --cov=ergon_core --cov=ergon_builtins --cov-report=xml:coverage.xml -> 933 passed, 1 skipped
• uv run ruff check ergon_core ergon_builtins ergon_cli ergon_infra ergon_ingestion tests -> passed
• uv run ruff format --check ergon_core ergon_builtins ergon_cli ergon_infra ergon_ingestion tests -> passed
• uv run ty check ergon_core/ergon_core ergon_builtins/ergon_builtins ergon_cli/ergon_cli ergon_infra ergon_ingestion/ergon_ingestion -> passed
• uv run slopcop --warn-only ergon_core/ergon_core ergon_builtins/ergon_builtins ergon_cli/ergon_cli ergon_infra ergon_ingestion/ergon_ingestion -> passed with warnings only
• uv run python -m ergon_core.core.application.testing.suppression_budget -> passed
• uv run xenon --max-absolute F --max-modules E --max-average C ergon_core ergon_builtins ergon_cli ergon_infra ergon_ingestion -> passed
• pnpm -C ergon-dashboard run generate:contracts:rest -> passed
• pnpm -C ergon-dashboard run generate:contracts -> passed
• pnpm -C ergon-dashboard run lint -> passed
• pnpm -C ergon-dashboard run typecheck -> passed
• pnpm -C ergon-dashboard run test:unit -> 135 passed

Notes

• Broader tests/examples -q still includes existing failures outside this stack; the new TRL example import test passes in isolation and with the focused stack suite.

[github-actions]

E2E smoke — minif2f

Screenshots pushed to screenshots/pr-132.

[github-actions]

E2E smoke — swebench-verified

Screenshots pushed to screenshots/pr-132.

[github-actions]

E2E smoke — researchrubrics

Screenshots pushed to screenshots/pr-132.