[biswapanda]

Summary

• Allow explicit renderer-based evaluation without changing the default chat evaluation path.
• Add a backend-neutral LoRA administration hook and propagate an optional returned adapter model through filesystem weight updates before notifying observers.
• Preserve native LoRA loading behavior and its model name.
• Require complete routed-expert branches when replay is configured, reject incomplete rollouts early, and stop after the configured rejection budget.
• Allow explicitly trusted custom VLM checkpoints through the shared trainer configuration.
• Check whether NIXL is configured before accessing NIXL-specific model fields, allowing aggregated workers to initialize with the tested vLLM RC.
• Resolve the image adapter only when a batch contains multimodal references, allowing text-only training without an image adapter while preserving image validation and caching.

Why

These are shared RL interfaces and validation rules, not backend-specific discovery or worker-management behavior. Evaluation needs an explicit request-construction mode; filesystem weight updates need a backend-neutral way to report the active adapter model; replay requires complete branch metadata before training; and explicitly trusted custom VLM checkpoints must pass configuration validation.

Aggregated workers should not inspect removed NIXL-only configuration fields when no NIXL connector is active. Text-only composite-model training should not require an image adapter at startup; image-bearing batches still resolve and validate the adapter before materialization.

Validation

• 162 focused inference, client, rollout, algorithm, and configuration tests passed against this branch after the two startup corrections were added.
• Five source-block checks passed for deferred adapter selection, text-only execution, adapter caching, and rejected unsupported/misconfigured image batches. These are source-block checks, not full trainer tests.
• The watcher regression passed in the SSH dev-pod runtime with GPU visibility disabled.
• The composed shared-plus-Dynamo stack passed 197 focused tests. Its production code exactly matches the original integration merged with current main; only a native LoRA regression test was added.
• Ruff and diff whitespace checks passed.
• The broader recipe-loading suite was not green in the local environment: 38 cases failed with environment-plugin/configuration issues. Those cases were excluded from the focused result above.
• The composed Nemotron + Dynamo stack completed three AutomationBench optimizer updates with filesystem LoRA, TP4 inference, strict routed-expert replay, confirmed v0-v3 deployed adapters, and all process exit codes 0. It used these exact two production corrections, additional Nemotron conversion/model support, and feat(dynamo): support python based dynamo vLLM workers administration #3629; this is composed-stack evidence, not standalone native acceptance of this PR head.
• That regression had 50% truncation in the first batch and token-level mismatch KL maxima up to 0.784544. It verifies loop/adapter/replay plumbing, not convergence or universally low KL. No new GPU run was launched for this publication.

Integration order

This PR extracts backend-independent support from #3629. Both target upstream main; merge this prerequisite first, then #3629.