Fix Dragon inference with local loader processes

[tpn]

Dragon xScan inference with --num-workers 1 stalls before READY because the local Python spawn child receives a Dragon-patched queue and cannot finish bootstrap. Keep Python multiprocessing unpatched when starting native Dragon workers, so their local DataLoader children use ordinary spawn queues while native Dragon coordination remains intact.

Validated on x86-64 B200 and two-node aarch64 GB200 with two GPU workers and three repeated rounds. Positive-loader inference and the default loader path complete; scientific outputs match the original candidate's working loader-zero baseline exactly, loader children persist across rounds, and shutdown leaves no owned processes. Separate fix wheels preserve the original candidate evidence. All 35 focused lifecycle tests, including a real Dragon spawn-queue regression, and pre-commit checks pass.

[coderabbitai]

Important

Review skipped

Auto reviews are limited based on label configuration.

🏷️ Required labels (at least one) (1)

• ai-review

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository: NVIDIA/cuPhoton/.coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 68811c6b-a997-4c92-a0f6-8f9e803c4de3

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Autopilot is currently an internal CodeRabbit preview.

---

_{Comment @coderabbitai help to get the list of available commands.}

[melo-gonzo]

LGTM. All 35 lifecycle tests passed with Dragon 0.14.2, and the new spawn-queue regression fails on the base revision. The environment overlay preserves inherited variables and leaves the coordinator environment unchanged. Multi-node B200/GB200 validation was not independently repeated.