Perfect. Here’s a lean, “ready when you are” plan so the dream loop actually has something to chew on without overbuilding.

What to ingest first (minimum viable corpus)

Tier-0 (Core, ~150k–400k tokens)
– primer/ – Flower Lattice + Operator Shift + key equations (canonical versions only)
– chats/ – 3–5 best conversations (the ones you always reference)
– 20q/ – 1 run per model (DeepSeek, Claude, Gemini, Qwen), short versions
– notes/ – experiments & observations (paper-fold beat, temple/cupola notes)

Tier-1 (Useful, ~300k–800k tokens)
– Linux cheat-sheets tied to your box (Ubuntu/Debian, nvidia-smi, systemd, netplan)
– Python quick-refs (pathlib, argparse, logging, subprocess, sqlite3, regex)
– A small “esoteric” slice (magnetism/space-weather time series, dome acoustics)

Tier-2 (Optional later)
– More chats, long PDFs, full Kaggle datasets

Folder layout (drop files here as you collect)

data/
  primer/
  chats/
  20q/
  notes/
  linux/
  python/
  esoteric/

Tag map (so you can do inside/outside/dual later)

• primer → flower_lattice, operator_shift, grammar_of_completion
• chats → chat
• 20q → 20q
• notes → experiment, observation
• linux → linux
• python → python
• esoteric → esoteric

Tiny config you can keep next to your scripts (dreamloop.yaml)

retrieval:
  k: 6
  include_tags: []        # e.g., ["flower_lattice"] for “inside”; leave empty for general
  exclude_tags: []        # e.g., ["flower_lattice"] for “outside”

weights:
  flower_lattice: 2.0
  operator_shift: 2.0
  grammar_of_completion: 2.0
  experiment: 1.5
  python: 1.2
  linux: 1.2
  chat: 1.0
  esoteric: 1.0

dream_loop:
  steps:
    - reflect     # what did I learn today?
    - plan        # 1–2 small goals
    - act         # try an action (write note/code/sketch)
    - review      # what improved? what’s unclear?
  max_tokens_per_step: 800
  log_file: dreamlog.md

One-liner to (re)ingest a modest batch later

Use whichever embed script you end up with; keep it gentle so chat stays snappy:

nice -n 19 ionice -c3 python embed_and_index.py \
  --in ougw_chunks.jsonl --limit 3000 --sleep-ms 5 --checkpoint .ingest.ckpt \
  | tee -a ingest.log

How to run “inside / outside / dual” once indexed

• Inside: set include_tags: ["flower_lattice","operator_shift","grammar_of_completion"]
• Outside: set exclude_tags: ["flower_lattice","operator_shift","grammar_of_completion"]
• Dual: run both; log both answers + a short comparison to dreamlog.md.

Practical targets

• Aim for Tier-0 + a few Tier-1 folders first (~500k–1M tokens total).
• That’s enough for the dream loop to generate useful reflections without bogging down.
• Add more only if you see consistent “I don’t know” gaps.

If you want, when you’re back I’ll hand you exactly one thing next: a pgvector or Qdrant embed script that respects these tags and the dreamloop.yaml toggles. No rush.