ci: add HF Space keep-alive and uptime-monitor workflows; docs

server: serve showcase HTML at / so HF Space embed mirrors *.hf.space/showcase (move root JSON to /info)

docs: update demo video link to https://youtu.be/aBnQ33q9MMw across README, BLOG, and showcase header

blog: inline ENV_DEEP/ENV_SHALLOW/TASKS_SHALLOW into BLOG.md (deep first); voice: tour deep regime in scene04

showcase: add Deep regime section (Rounds 1+2) with 11-archetype table, per-checkpoint Round 1 numbers, v2/v3/v4 critic swap deltas, and per-task v2->v4 improvement table

docs: add TASKS_SHALLOW.md - row-by-row index of all 381 procedural scenarios

docs: add env reference for deep (11-task) and shallow (381-task) regimes; expand BLOG appendix

blog+readme: document the two prior deep-training rounds (legacy SB3 + hybrid Ollama/Groq PPO on 11 tasks); cite training code; add 3-panel deep-training chart

blog: replace mermaid fences with rendered PNGs (some viewers don't render mermaid); tighten legacy chart y-axis

blog: embed legacy + LLM training graphs and 3 mermaid pipeline DAGs

blog: add BLOG.md (story + novelties + training proofs); README: link YouTube + blog; voice: auto-recorder + scene06 fix

showcase(results): add evidence-of-learning section + scene04 voiceover walks legacy baseline vs PPO Kaggle, KL/loss convergence, per-category Δ

showcase(slack): wrap long lines in code snippets so they don't overflow the card

showcase: add Slack-noise novelty section + DAG; voiceover scene03 mentions coworker Slack; 4:58 total under 5min cap

fix(showcase): visible task grid + reward-sign explainer; fix(dashboard): rewards stretch glitch + dual legacy/ppo dashboards with sourced-file banners

feat(showcase): /showcase page + precomputed bundle + READMEs + video script

feat(train): show current task_id in heartbeat lines so round-robin coverage is visible at a glance

fix(collector): persist task cursor across PPO updates - was resetting to 0 each update so only first IC_ROLLOUTS tasks were ever trained on; now round-robins full shard (3 rollouts x 60 updates = 180 visits over 127-task shard = full coverage); also log first/last 5 task ids at startup so coverage is visible

fix(train_lib): wrap unsloth for_training/for_inference in try/except — peft+unsloth compat bug deletes _flag_for_generation; falls back to model.train()/eval() (correct, slightly slower); add IC_DISABLE_UNSLOTH escape hatch

feat: hand-written 3 Kaggle ipynbs with unsloth install attempt; train_lib disables actor grad-checkpoint + filters Phi3 warning spam (2x faster + clean logs)

fix: drop trust_remote_code so transformers uses native Phi-3/Qwen3 (custom modeling code is incompatible with t>=4.48 DynamicCache API)

fix(kaggle): bump transformers to >=4.51 for Qwen3 (DeepSeek-R1-Qwen3-8B critic)

fix(kaggle): chdir out of WORK before rmtree so re-run git clone succeeds

fix(kaggle): always wipe + re-clone repo so HF_TOKEN-optional fix is picked up

feat(train): make HF_TOKEN optional + continuous live ETA heartbeat during rollouts

feat(kaggle): use user-pinned mount paths Phi-3.5-mini (actor) + DeepSeek-R1-0528-Qwen3-8B (critic); train_lib: trust_remote_code + adaptive LoRA target_modules for fused-qkv archs

fix(kaggle): switch to Phi-3 (MIT, no access gate) since Llama license requests are disabled

fix(kaggle): use Kaggle Models (Llama-3.2-1B + Llama-3.1-8B) instead of HF download to avoid filling /kaggle/working

fix(kaggle): download Qwen actor+critic from HF Hub instead of Kaggle Models

feat(kaggle): 3 sharded training notebooks + local-critic mode + LoRA merge

feat(phases): realistic 3x12 configs for phase2/3 launchers

fix(hfjob): bitsandbytes 0.45.5 to avoid triton.ops removal

fix(hfjob): pin transformers/peft/bnb stack — skip unsloth (gemma3n compile crash)

fix(actor): fallback to plain HF transformers when unsloth fails to import

fix(hfjob): install unsloth_zoo + datasets/sentencepiece deps that unsloth needs

fix(hfjob): use python3 interpreter (HF transformers image has no python alias)

feat(train): warm-start adapter loading + IC_TASK_MODE (curated/all/hard) + phase2/phase3 launchers

feat(train): live tqdm+ETA, per-update HF Hub checkpoint upload, HF Jobs entrypoint + launcher

fix(colab): silence Qwen max_length FutureWarning, real 120-update training run, validated task IDs

feat(phase8-10): saboteur+slack+replay, 381 scenarios, Colab Unsloth+Qwen72B critic, push helpers

Remove kube-sre-gym comparison from README

fix(dashboard): hoist unicode literals out of f-string expression for py3.11

v4: Groq critic, real AWS writes, +17 reward signals, dashboard polish

rename hybrid-ollama-bedrock -> ppo-v3-hybrid-ollama-bedrock so it sorts last and becomes dashboard default

fix: replace f-string with backslash escapes that broke dashboard_pages import

Hybrid Ollama+Bedrock actor-critic trainer + dashboard snapshot pipeline

Real training data + full AWS multi-service infra

fix(hf): copy dashboard_pages.py into image + avoid Py3.11 f-string backslash

fix: include dashboard pages in HF image

feat: actor-critic PPO trainer + multi-page analytics dashboard