Unify trainer reconciliation and validate records at the KV boundary

[kailash109]

Summary

• Shared, scoped, and local deployments now scale trainers through one reconciler. Scoped runs use the same path with limits of 0–1 instances and one model per instance, instead of a separate reconciler or parameterization-based routing.
• Definition selection is one alias map: an explicit definition id, or the first recipe registered for a Hugging Face model name. Sampling sessions require an engine definition id.
• Durable session, model, placement, and sampling records are validated once at the KV boundary, and legacy untyped sampling-session records are no longer accepted.
• Package metadata, distribution checks, and publish workflows prepare the modal-spindle PyPI release. deploy_models.sh includes the 128k Qwen LoRA recipe and deploys through uv on Python 3.12.

Test plan

• Run the CPU pytest suite
• Create a model and confirm trainer_demand is written and reconciliation scales trainers for that definition
• Request a model by Hugging Face name and by explicit definition id, and confirm the first registered recipe is the name default
• Run uv build and scripts/check_distribution.py

Made with Cursor