Add reproducible Cosmos3 Nano and Edge fine-tuning

[ramanathan831]

Summary

Add repository-owned Cosmos3 Nano and Edge fine-tuning support, including:

• Dataset-neutral video-QA recipes, deterministic data overrides, resumable batching, and worker prefetch.
• LoRA training, Qwen patch-embedding compatibility, token metrics, epoch checkpointing, and structured lifecycle reporting.
• Repeated-video validation caching, batched vision attention, and opt-in gradient-spike rollback.
• Edge attention/mRoPE compatibility, portable FP8 cache decoding, native VLM checkpoint export, and image/source provenance.
• Non-root container support and the validated pre-commit dependency bootstrap.

Cleanup and scope

• Replace the recipe re-export and dataset-specific monolith with direct registration of four generic Nano/Edge conversation/reasoning recipes.
• Move reusable video processing/collation into the existing dataflow roles module and media-grouped sharding into the existing distributor module.
• Use the Framework-native reasoning dataset and lifecycle callback directly; training has no external toolkit runtime dependency. Their integration is promoted from the dependent runtime PR.
• Retain three generic examples because their model profiles and training schedules differ. Generalize the launcher and environment names without changing those schedules.
• Retain the rollback guard, A100 compatibility helper, offline VLM exporter, and regression coverage: these provide independent capabilities.
• Preserve strict provenance for explicitly identified release images while allowing ordinary builds to report unverified source identity. No CI gates, configurations, or pins changed in this cleanup.

This is structural consolidation, not a blanket file deletion: the current diff has 82 changed files, including 23 new files. Shared-module reuse and moving the native dataset into this parent add edits to existing files; the dependent runtime PR shrinks from 102 changed files to 80.

History and dependent PRs

Supersedes #164, which this account cannot reopen after its maintainer closure. The existing source branch is retained.

The initial 58-commit source history was consolidated to 14 logical commits with an identical final tree. The subsequently requested implementation cleanup adds three focused commits: native video SFT consolidation, generic examples, and default/release image provenance. The current tree therefore intentionally differs from the earlier history-only snapshot.

The dependent PRs are restacked, not replaced:

• Runtime: feat: add native Cosmos workflow runtimes ramanathan831/cosmos-framework#3
• Skills: feat(skills): integrate Cosmos workflow capabilities ramanathan831/cosmos-framework#1

Validation at e4481fc

• 103 targeted regression tests passed on this parent, including callbacks, TOML configuration, video recipes, resume, validation caching, loss metrics, attention/mRoPE compatibility, LoRA, export, and Dockerfile/provenance contracts.
• The four registered recipe configurations match the previous versions. The five relocated classes have equivalent executable ASTs (apart from the distributor's generalized dataset type annotation).
• All three generic example TOMLs load through the real Framework configuration loader.
• Both original pre-commit configurations pass locally on all three branches.
• The restacked dependent branches pass 118 native runtime tests, 557 skills CPU-contract tests under Python 3.11, and strict version stamps.
• Fresh hosted fork pre-commit and CPU-contract checks passed on the restacked heads. Upstream workflows still require NVIDIA maintainer approval; these fork checks do not establish an upstream CI pass.
• Previous A100/H100 functional validation is recorded separately. GPU flows and container builds were not rerun for this structural cleanup, and the unchanged H200-specific cases remain unverified without the upstream hardware. The existing Docker Build job is disabled by its upstream if: false condition.

This replacement remains a draft pending upstream CI and review.