openspec: S24 sampler island, trainer start/stop, decoupled PPO, identity tiering

[michaellchung]

Summary

Adds the OpenSpec change rl-trainer-cold-start-cost (proposal, design, tasks, spec delta). Spec only; no code changes.

Goal: lower the start cost of a trainer island, so that a trainer can start and stop often (S24 sampler island work).

• L0: seven start phase events; the scheduler reads the start time T_s from them.
• L2: Miles import chain fix (light actors stop importing miles.backends.megatron_utils; lazy modelopt import), setup reads the image manifest instead of importing peft/sglang, peft 0.20.0 in the ports image, replace the broken huggingface-cli download prefetch and record failures.
• L1.1: measure a separate trainer island in two steps: Qwen3.5-4B on 2 cards (TP2) first, then decide on Qwen3.8-27B on 8 cards.
• L8: SGLang memory snapshot for samplers only, gated on a GPU check that it works with RL weight updates.

All effort and GPU cost numbers in tasks.md are estimates. Nothing has run on GPU for this change yet.

Validation

• openspec validate rl-trainer-cold-start-cost: valid.

🤖 Generated with Claude Code

[michaellchung]

Superseded by the stacked draft PRs #262 (cold start) → #263 (off-policy correction) → #264 (identity tiering) → #265 (start/stop) → #266 (TB2/27B run support). Each implementation PR carries its own openspec change. Keeping this one open for reference until the stack is reviewed.