chore: ignore figure thumbnails; tidy gitignore

Regroup sections; add data/pages/*/thumbs/ (regenerable via
scripts/render_figure_thumbs.py, not committed).

Author: NorthernLightx

.gitignore

@@ -12,13 +12,17 @@ env/
 dist/
 build/
 
-# Testing
+# Testing & coverage
 .pytest_cache/
 .coverage
+.coverage.*
+coverage.xml
 htmlcov/
 .mypy_cache/
 .ruff_cache/
 .playwright-mcp/
+.tox/
+__pypackages__/
 
 # Editors
 .idea/
@@ -30,43 +34,15 @@ htmlcov/
 .DS_Store
 Thumbs.db
 
-# Env
+# Env & secrets
 .env
 .env.local
 .env.*.local
+*.pem
+*.key
+secrets/
 
-# Project artefacts
-data/papers/
-data/golden/raw/
-# Pre-review WIP; promoted candidates land in tracked data/golden/<set>.yaml.
-data/golden/_candidates/
-data/eval/runs/
-data/figures/
-# Page PNGs ignored except the committed 2604.* demo corpus (baked into the image).
-data/pages/*
-!data/pages/.gitkeep
-!data/pages/2604.*
-# Regenerable DPI render scratch.
-data/pages_dpi*/
-# Transient portalocker sentinel (the qdrant_local/ snapshot itself is committed).
-qdrant_local/.lock
-# Only the papers.txt manifest is committed; PDFs fetched on demand.
-data/curated_demo/*
-!data/curated_demo/papers.txt
-# Fetched benchmark data; only the derived golden YAML is committed.
-data/mmlongbench/
-# Fetched BRIGHT corpus; regenerated by dci_prepare_bright.py.
-data/dci/
-logs/
-*.log
-.langfuse/
-qdrant_storage/
-postgres_data/
-
-# Personal cost kill switch, not part of the deployed app
-deploy/cost-guard/
-
-# Coding-agent / LLM-tooling local state and agent-specific docs
+# Coding-agent / LLM-tooling local state
 .claude/
 .cursor/
 .windsurfrules
@@ -75,59 +51,68 @@ CLAUDE.md
 GEMINI.md
 AGENTS.md
 docs/superpowers/
-# Local research notes.
-docs/research/
 
-# Git worktrees (isolated dev workspaces)
+# Git worktrees
 .worktrees/
 
-# Coverage / packaging leftovers
-.coverage.*
-coverage.xml
-.tox/
-__pypackages__/
-
-# Internal rolling-state notes (per-author project tracking; not for public repo)
-STATUS.md
-docs/mvp-plan.md
-
-# Terraform local state (.terraform.lock.hcl IS committed for provider pinning)
+# Terraform local state (.terraform.lock.hcl is committed)
 **/.terraform/
 *.tfstate
 *.tfstate.*
 *.tfplan
 
-# Secrets
-*.pem
-*.key
-secrets/
-data/eval/ingestion/overlays/**/*.png
-data/eval/ingestion/vlm-spike/*.png
-
-# Regenerable inspection/study artefacts (local-only).
-data/ingestion_study/
-data/tiny_figures_inspect/
-data/unlabeled_inspect/
-# Single-paper smoke-test corpora.
-data/papers_one/
-data/papers_*_smoke/
-
-# Local human gold-review tool + its saved verdicts (Bet 1 eval audit).
+# Local notes & tooling, not deployed
+docs/research/
+STATUS.md
+docs/mvp-plan.md
+deploy/cost-guard/
 tools/gold_review/
-data/eval/audit/
-
-# Isolated MinerU extractor venv + its parse output (ADR 0025).
 tools/_mineru_venv/
 tools/_mineru_out/
-# Local agent/workflow run artifacts.
+
+# Runtime & output
+logs/
+*.log
+.langfuse/
+qdrant_storage/
+postgres_data/
 output/
 
-# Local debug scratch at repo root.
+# Local debug scratch at repo root
 /*.png
 /*.jpeg
 /.tmp_*.json
 /figs*.json
 /*_tmp.json
+
+# Corpus & data: fetched or generated locally
+data/papers/
+data/papers_one/
+data/papers_*_smoke/
+data/golden/raw/
+data/golden/_candidates/
+data/eval/runs/
+data/eval/audit/
+data/eval/ingestion/overlays/**/*.png
+data/eval/ingestion/vlm-spike/*.png
+data/figures/
+data/pages_dpi*/
+# regenerable figure thumbnails
+data/pages/*/thumbs/
+data/mmlongbench/
+data/dci/
+data/ingestion_study/
+data/tiny_figures_inspect/
+data/unlabeled_inspect/
+# page PNGs: keep the committed demo subset
+data/pages/*
+!data/pages/.gitkeep
+!data/pages/2604.*
+# only the papers.txt manifest is committed
+data/curated_demo/*
+!data/curated_demo/papers.txt
+
+# Local Qdrant snapshot: the dir is committed, these are not
+qdrant_local/.lock
 qdrant_local_old/
-# Too large to commit; baked into the image (ADR 0028).
 qdrant_local/collection/rag_corpus_visual/