[Train 3/3 — after #78] Schedule local SFT across the Grid

[deehw]

What

• add grid train submit-sft for a portable config + verified SFT dataset
• reuse the durable task/Git plane; no second scheduler
• advertise explicit train-mlx and train-torch worker capabilities
• claim only on dependency-compatible nodes and return adapter artifacts in the result commit
• recover a lost machine through the existing lease/reclaim path
• give Train bounded per-job queue/run clocks instead of the ordinary one-hour agent-task limit
• reject zero-exit false success unless a real non-empty adapter and matching SFT run record exist
• add a portable SHA-256 result manifest and grid train verify-result

Honest scope

• one trainer owns one job; this is node scheduling/failover, not cross-machine DDP
• v1 restarts from immutable input after node loss rather than resuming optimizer state
• training capacity is opt-in and absent from the default worker list
• relay-first rollout; private relay companion PR required

Safety

• shell-free, allowlisted job spec
• isolated Python launch prevents a checkout from shadowing Grid's trainer package
• config/data/result paths are confined to the task worktree
• credentials require explicit passthrough
• fetched results reject missing, added, symlinked, changed, or metadata-mismatched adapter files

Real acceptance

• isolated self-hosted relay with production-like 120-second leases
• real mlx-community/SmolLM2-135M-Instruct one-iteration SFT
• task fbd1d906-5f9b-46d2-9e02-45ec9c59f843 completed on attempt 1
• result commit 4ac1e59098076e820010b467139814939d20187b fetched into a clean client
• 1,315,662-byte adapter independently verified from per-file SHA-256 hashes
• controlled lease-loss tests prove attempt 2 keeps the identical immutable input commit on a different compatible provider

Validation

• full public suite: 3,425 passed, 57 skipped, 7 deselected
• focused Train suite: 7 passed
• Ruff: clean
• wheel built and inspected for the Train CLI and worker modules
• private relay companion: final-base local + GitHub gates 326 passed each; shared tasks 129 passed; eligibility clocks 38 passed

Stack

Depends on #78, which depends on #77. Merge autonomous-ai/autonomous-grid-cli#22 before enabling training workers.