chore(spaces): add HF Space config for training [skip ci]

final submission

fix(sft): move max_seq_length + dataset_text_field to SFTTrainer

fix(training): follow unsloth's Qwen3 guide exactly

fix(training): upgrade base image to torch 2.6.0+cu126

feat(training): switch from LoRA to QLoRA per mentor recommendation

fix(training): pin torchao==0.5.0 after unsloth to fix torch.int1 error

fix(spaces): switch training Space to A10G Small, tune notebook for 24 GB

fix(spaces): add app_port: 8000 to env Space YAML

fix(training): upgrade to torch 2.5.1+cu124, restore unsloth for Qwen3

fix(training): drop unsloth, use bitsandbytes+PEFT for SFT

fix(docker): pin torchao==0.5.0 + transformers/trl/peft before unsloth install

fix(notebook): Cell 3 load_dataset per split + write JSON (was snapshot_download finding 0 .json files)

fix(docker): COPY README.md into build context (hatchling metadata requires it)

fix(spaces): orange→yellow (not a valid HF colorFrom value)

fix(spaces): inject HF Space YAML frontmatter into README on push

chore: add push_to_hf.sh — one-shot deploy to both HF Spaces

feat(training): A10G-optimised pipeline — auto train.py, Dockerfile.train, GH Action sync

parallelized the data generation for openai

perf(trajectory_gen): parallel workers + JSONL checkpoint for resume

fix(trajectory_gen): add --scenarios-dir flag for server-free generation

fix(submission): Dockerfile, wire-format fixes, scenario loading, real-scenario MockEnvClient

merge(branch-b): data pipeline + rewards + training — B1-B5, C1-C6

merge(branch-a): env core — FastAPI server, 11 tool routes, episode lifecycle, replay visualizer (A1–A5)

feat(training): Phase C6 — ablations, training curves, readme finalization

feat(training): Phase C5 — evaluation harness, baselines, plots, readme table

feat(training): Phase C4 — GRPO training, SFT warmstart, rollout, custom trainer

feat(training): Phase C3 — SFT trajectory generator, env clients, mock env

feat(rewards): Phase C2 — composite reward, replay verifier, frozen weights

feat(rewards): Phase C1 — all 9 reward components implemented

feat(data): Phase B5 — corpus instantiation, HF publish, annotations

feat(data): Phase B4 — 7 ScenarioFamilyGenerators with archetype loading

feat(data): Phase B3 — failure clustering and archetype extraction

feat(branch-b): B2 GitHub Actions log mining — gh-CLI scraper, anonymizer, throttle, cache

feat(branch-b): B1 public dataset ingest — DeFlaker, iDFlakies, FlakeFlagger, LogHub loaders + CLI

feat(branch-a): A5 replay visualizer — static HTML/JS viewer mounted under /viz

feat(branch-a): A3 episode lifecycle — budget enforcement, terminal handling, payload truncation, trace serialization

feat(branch-a): A2 tool implementations — route 11 handlers to scenario.tool_outputs with cost charging

refactor(branch-a): A1 — adopt openenv-core MCPEnvironment + create_app for canonical OpenEnv compliance

feat(branch-a): A1 server scaffold — FastAPI /reset /step /state /mcp + 11 stub tool handlers

feat(phase-0): foundation — uv project, schemas, mocks, manifest, CI

chore: import planning artifacts