[michaellchung]

Body:

Summary

Integrates the S25 work after #278 (s25-n71d) into one branch:
s25-e2v (integ4: G2/G3/G4, C1-C4, C3V, T-EVAL, E2V) + s25-c1b + s25-g1c + s25-n109.
89 non-merge commits. Some C1 commits appear twice (once from s25-c1, once
cherry-picked into s25-g1c); the patches are identical and the merge keeps one tree state.

Sampler islands and spot capacity (C1, C1b, E2V, G1C)

• --rl-sampler-spot: spot sampler islands with an on-demand trainer; several samplers per trainer; the trainer's rounds split evenly over the samplers.
• Reclaim handling: the reclaim signal becomes a marker inside the island; a watcher thread leaves at once; the signal handler never raises; engine server adapters ignore SIGINT so the island finishes its round and leaves PREEMPTED.
• A replaced sampler island (Miles and engine sampler) resumes from the next round and never overwrites a payload.
• Nebius spot: a host unit forwards the VM stop; the unit install works with a remote sky API server.
• codepack: lists executables and restores lost executable bits; fixed mtime so sky can zip; reuses a tree another launch unpacked.
• Miles buffer reader records passed-over stale groups (rl_buffer_stale_skipped).

Trainer islands (C2, C3, C3V, C4, N7.6b/c)

• verl trainer island over the sample buffer.
• Synced verl and Miles trainer islands feed sampler islands through merged releases; merged releases over S3/OSS object stores.
• yeto round-cut controller in the verl island; SCHEDULED cut + resume; non-strict load for frozen LoRA base weights; ChainedOptimizer support; cut coverage of dense tensors over EP ranks.

Stage 1 (GLM5.3-Next) rehearsal fixes (G1, G1C, G2-G4)

• Stage-1 plan: sampler.extra_args, trainer.miles.extra_env, task.data_local, AlgorithmSpec hash pin, RAY_ADDRESS export, trainer-only island without engines.
• glm5_next trains from a Megatron torch_dist checkpoint (raw mode); geometry from config.json.
• Re-quantizer: raw glm5_next export prefix, convert_to_hf name map, host copies from the Miles adapter.
• SGLang expert load report (post-load check 1); general tp/dp topology check.
• Token-diversity test reward; engine sampler rewards see token ids.

Agentic RL (N10.9)

• verl Harbor path: records proxy over verl's rollout client, path-secret recorder, agentic entry, min/max_global_steps per sample.
• Miles Harbor path with the records proxy as recorder.
• qwen3_2507 chat template with token alignment tests; outcome rules: a malformed reply or a context overflow is a trained 0, not ABORTED.

Evaluation suite (T-EVAL)

• tools/eval (list/run/judge/report): capabilities E1-E5 x Miles/verl on Modal A10G, Qwen3-0.6B.

Merge notes

• s25-g1c conflicts: all conflicting hunks came from the patch-identical C1 commits and a4b9962 (= bd58b18); this side was kept because it also carries C1b/E2V changes. The SGLang topology hunk keeps the general rule, which also admits the rehearsal's tp1.
• s25-n109 merged without conflicts; its files are disjoint from the other branches. Completion usage checked against the G1C prompt_token_ids field.

Tests

• Remote full suite (yeto-cpu-test): 6628 passed / 0 failed / 94 skipped (SESSION24/logs/t-final25-full2.log).
  The first run had 1 failure in test_delta_protocol (unchanged file, timing); 5 single re-runs passed, the second full run passed.
• Tokenizer tests with HF_HUB_CACHE: 19 files, 375 passed + test_records_proxy 21 passed (SESSION25/logs/t-final25-tokenizer.log).
• yeto-stage1 plan examples/stage1/glm53-flash-ppu.yaml: rc=0; 2 known GAPs (PPU trainer image N0.2, docker device args N0.2b).
• python -m tools.eval list: lists 10 cells.
• GPU evidence per branch: C1b s25c1b-d, G1C g1c-1, E2V e2-verl-e2v-b, N10.9 runs c/d, T-EVAL 9/10 cells (see SESSION25/GOAL.md).

Not verified

• The merged branch itself was not run on GPU; GPU evidence comes from the source branches.
• Real cloud reclaim (C1b used nebius compute instance stop).
• Agentic RL reward increase (N10.9 proves the chain only).
• PPU/NPU runs; qwen38 and fork-template tests skipped (no tokenizer / fork checkout).

🤖 Generated with Claude Code