refactor(observability): decouple trainers from WandB

[leviking98z-rgb]

Summary

• Add a provider-neutral Observer protocol, NullObserver, factory, checkpoint-state helper, and phase instrumentation under unirl.observability.
• Make BaseTrainer, every loop program, and all trainer variants emit metrics, media, progress, and lifecycle events through self.observer; trainers no longer import or name the WandB adapter.
• Preserve existing report_to_wandb recipes while adding explicit logging.provider: none|wandb and generic logging.enabled selection.
• Preserve disabled-run behavior without loading the WandB adapter: console progress still prints, memory-monitor step boundaries still fire, and phase timing remains injected at the rollout-step boundary.
• Extend the framework guard so the Observer surface remains stable and trainer code cannot reintroduce direct WandB imports, wandb_logger, or backend-named lifecycle hooks.

This is a boundary-stability PR, not a LOC-reduction PR: 630 insertions / 364 deletions (net +266), including the new contract, null implementation, static guard, and documentation. It removes the duplicated timing/progress policy from the concrete adapter and makes another telemetry backend possible without changing trainer programs.

This is stacked on refactor/reward-ownership (#31 in this fork), not on main.

Related Issue

N/A

Test Plan

• pre-commit run --all-files — passed, including all 120 recipe compositions, target resolution, Ruff, architecture boundaries, experimental boundaries, and framework contracts.
• python -m compileall -q unirl/observability unirl/trainer unirl/utils/wandb_logger.py unirl/utils/memory_monitor.py unirl/utils/media.py unirl/types/media_preview.py lint/check_framework_contracts.py — passed.
• python lint/check_framework_contracts.py — passed; reports one Observer contract and verifies no trainer-facing WandB dependency remains.
• python lint/check_architecture_boundaries.py — passed (476 files / 4,279 internal import edges).
• python lint/check_experimental_boundaries.py — passed.
• One-off stdlib CPU harnesses exercised structural Observer conformance, old/new checkpoint run-id state, disabled factory selection, invalid-provider rejection, enabled-adapter argument forwarding, phase-time injection across consecutive steps, console progress, and memory-monitor step boundaries — passed.
• Not run: GPU training or live WandB upload. Metric extraction, WandB SDK calls, and training numerics are unchanged; this PR changes their ownership and call boundary.

Compatibility / Risk

• Existing logging.report_to_wandb, project, tags, entity, directory, and media fields remain accepted with the same enabled behavior.
• New explicit configuration uses provider: wandb plus enabled: true; provider: none selects the dependency-light null observer. Unknown providers fail before training starts.
• Checkpoints now write observer_run_id and continue writing wandb_run_id; resume accepts either field, preferring the provider-neutral one. Older and newer checkouts can therefore exchange trainer state.
• BaseTrainer.wandb_logger, _init_wandb, _finish_wandb, and _loop_wandb_extra are replaced by provider-neutral names. Out-of-tree trainer subclasses using those internal seams must migrate to observer, _init_observability, _finish_observability, and _loop_observability_metadata.
• Phase and memory instrumentation still wrap collaborators lazily on the first training step. No recipe, model, reward, dataset, tensor wire format, or GPU placement changes.

Reviewer Notes

• Suggested review order: unirl/observability/api.py, factory.py, trainer/base.py, trainer/program.py, then check_observer_contract.
• Duplicate-work check: no open official PR introduces a provider-neutral observability contract. Official refactor(trainer): extract the sync-loop template into BaseTrainer Tencent-Hunyuan/UniRL#282 also edits BaseTrainer/AR/diffusion to extract a synchronous loop, but this stack already centralizes all loop families in refactor(trainer): separate lifecycle and loop programs #25 and this PR addresses a different boundary. Official fix(trainer): preserve async diffusion teardown cause Tencent-Hunyuan/UniRL#239 and feat(async-ar): full-async dual control for continuous rollout Tencent-Hunyuan/UniRL#213 touch async teardown/trainer files but do not abstract metrics or media providers; textual rebase conflicts are possible, semantic duplication is not.
• AI assistance was used. The complete diff, compatibility paths, and CPU/static validation evidence were reviewed before publication.

Checklist

• I reviewed the changed code and removed unrelated/generated artifacts.
• I updated tests, docs, and configs where needed, or explained why not.