sync: pull axentx-master 7d05ef5 (workers 1→4 per Space, push 2×)

sync: pull axentx-master c14673b — twin harvest worker

feat(burst-but-dont-die): adaptive auto-scaler + tighter discoverer cycle

fix(oom-permanent): memory guard + redis cap + heavy-task gating

fix(oom): stagger cron offsets + STREAM=1 — minute=0 burst was OOM trigger

feat(regress-cron): daily 11:00 UTC regression-test on both Space + anchor

feat(round12-tier2-regress): GSPO + CodeScaler stubs + 10-step regression suite

feat(round11-anchor-devchain): full dev-chain on anchor + axentx target verified

fix(low-mem): prune daemon stack — Space OOM after Round 9+10 added too much

feat(round10-continuous): always-on multi-source crawler + tight worker idle

feat(round9-anchor-parity): full Round 5+6+7+8 cron loop on OCI A1.Flex

feat(round8-trillion): 207 sources + streaming worker + enrich pipeline + dual-mode mirror

feat(round7-tier1): 4 frontier-2026 techniques (low effort, high impact)

feat(round7-scale): +90 datasets + aggressive HF harvester + unified llm-ladder

feat(pro-bridges): hf-inference + zero-gpu + submit-hf-job

fix(start): LOW_MEM mode for cpu-basic — Space dead 7h on OOM

feat(bridges): add openrouter + gemini bridges (v2 ladder needed them)

feat(v2-round6): TruthRL + Validator-RLVR + LoraHub + Self-Refine + Letta + 9-cluster configs

feat(v2-round5): sustainability loops + 2026 techniques

feat(bulk-mirror): 100+ massive datasets + 4-worker claim-queue coordination

feat(always-on): Ollama local fallback for ALL provider ladders

feat(v2): ambitious targets + 5 free-technique implementations

feat(v2): Phase A complete build infrastructure ready to execute

docs(v2-research): persist 16-stream research corpus + Phase A runbook

feat(sanitize): training data leak filter — drop rows w/ FS paths, LLM-provider tags, secrets, PII

fix(audit-2026-04-29): throttle gh-actions-ticker 60s→600s + harden /data symlink recreation

fix(boot): disable train-ready-pusher at Space boot — caused RUNTIME_ERROR; script preserved for manual launch

feat(train-ready): single-file pusher train-ready/latest.jsonl.gz for CDN-only training fetches — eliminates HF API listing during Lightning/Kaggle training

fix(mirror): write to batches/mirror-merged with clean {prompt,response} schema only — drops source/ts extras + raw-mirrors/enriched paths that broke v1 training pyarrow cast

fix: OOM 16Gi — mirror row-group stream + watchdog tier-2/3 kills

fix: orchestrate dev chain — REPO_CONTEXT unbound + provider 402

fix: Lightning Studio kwargs + Kaggle title-slug mismatch

fix: llm-burst — tolerate /data Errno 5 with retry + /tmp overflow

fix: dataset-mirror — stamps in /tmp + I/O error tolerance

fix: stamps — retry sources that previously kept=0 (extractor bug)

fix: dataset-mirror schema extractor — handle messages/ShareGPT/DPO formats

feat: Surrogate-1 model = PRIVATE on HF Hub (proprietary axentx asset)

fix: lightning use Python SDK (not deprecated 'lightning run app' CLI)

fix: dataset-mirror — enrich + dedup before upload (per user feedback)

feat: dataset-mirror — 100x faster ingest by bulk-cloning community SFT mixes

trigger: rebuild for LIGHTNING_* secrets

fix: lightning-trainer env name LIGHTNING_USER_KEY -> LIGHTNING_API_KEY

feat: 3-track training pipeline — Kaggle T4 + Lightning H200 + (Modal pending)

fix: kaggle slug date-stamped to dodge 409 Conflict

trigger: force Space rebuild to pick up KAGGLE_USERNAME=longlum secret

feat: +3 HF Inference providers — Llama 3.3 70B + Qwen-Coder 32B + Mistral 24B

fix: kaggle-trainer 7B -> Qwen3-Coder-30B-A3B (latest gen, MoE, T4x2 fit)

trigger: kaggle-trainer on every boot + cron 90-min (was 6-hr)

feat: kaggle-trainer + llm-burst x3 — TRAIN now, MAX synthesis

feat: +800 LLM-expanded SDLC keywords across 10 roles