HDGL UNIFIED ARCHITECTURE — verified state as of 2026-07-11 (night build)
========================================================================

ONE HDGL IMAGE BOOTS THE WHOLE STACK (/init order):
  1. phi_pool settles substrate to phi           [core, ticks forever]
  2. Temple pane : HDGLAgent.HC + PaneBus.HC      [ExePutS live-mutation]
  3. Linux sidecar: emerge-sidecar.sh             [stable LLM brain]
  4. Glove (Phase2): glove-train.sh               [Trailblaze critic learns]
  BUS: /lattice/slots (Delta) + /lattice/bus/{to,from}_bot (HolyC exchange)

TWO MUTABILITIES, SEPARATED (the key design):
  - brain never rewrites its own runtime  (stable, boring)
  - Temple pane is the ONLY mutable surface, mutates live via ExePutS
  Each system does the thing it is strong at.

THREE SHELLS — borrowed now, discarded by the roadmap (nautilus, not molt):
  SHELL 1 LLM : LM Studio/Ollama -> discard when glyph-model settles alone.
                JETTISON TRIGGER (measurable): glove routing-quality plateaus
                high WITHOUT the stock LLM's logits -> it learned the shape.
  SHELL 2 GPU : traditional CUDA -> discard when analog substrate IS the
                parallel compute (onion-shell HW). Custom no-cuBLAS kernels
                (tb_gguf_dequant.cu analog dispatch, conscious_fused_engine.cu
                warp-fusion) are the bridge; verified present, GPU-untested here.
  SHELL 3 OS  : Linux sidecar    -> discard when Temple-native bot stands.

VERIFIED BY EXECUTION THIS SESSION (not asserted — see PHASE2_VERIFIED.txt):
  [✓] phi-lattice settles (M_U=0.9999) + phi_fold32 avalanche 50% (zero-SHA)
  [✓] phi_stream seal/unseal + tamper + epoch forward-secrecy PASS
  [✓] critic (glove) LEARNS: V(good)-V(bad)=0.92; save/load persists exactly
  [✓] full CPU engine builds + runs: synthetic forward pass, HDGL routing,
      8 tokens @ 222 tok/s, --hdgl-alpha honored (0.20 vs 0.50)
  [✓] GGUF dequant F16/BF16/Q8_0/Q4_0/Q4_K all PASS with exact known values
  [✓] glove-train.sh uses ONLY real verified engine flags (corrected from a
      first draft that invented flags — caught + fixed)

STILL UNVERIFIED (honest, environment-bound):
  [ ] correct TOKENS vs a real Qwen3 GGUF — needs the 2.1GB file; HuggingFace
      + josefkulovany.com + forum.zchg.org all blocked by sandbox allowlist.
      Machinery proven; specific weights not loaded. Have-the-file gap.
  [ ] HolyC pieces (HDGLAgent.HC, PaneBus.HC) compile-clean on real TempleOS —
      symbols verified present; no HolyC compiler in this Linux container.
  [ ] GPU/CUDA path — needs NVIDIA driver in rootfs + a GPU; kernels present.
  [ ] end-to-end on real hardware (the "final slot-population" class of item).

FILES (all in outputs/):
  boot/substrate : hdgl-substrate-boot-merged.zip, init.expanded/.sidecar,
                   run-expanded.sh, EXPANSION.txt, SWAP_MANIFEST.txt
  Temple-native  : HDGLAgent.HC (self-mod agent), PaneBus.HC (bus receiver)
  sidecar        : emerge-sidecar.sh (brain), glove-train.sh (Phase 2)
  engine         : tb_infer_cpu_verified (runs), build_cpu_verified.sh,
                   critic_learn_test.c
  assessments    : PRE_DREAM_ASSESSMENT.md, PHASE2_VERIFIED.txt, this file

NEXT GROUND (for when you're back):
  - drop a real Qwen3 GGUF at /lattice/models -> close the tokens gap in one run
  - allowlist josefkulovany.com + forum.zchg.org -> fold in the onion-shell
    hardware corollary (SHELL 2's native successor) + the demo ingredients
  - Phase 3: native glyph-weight model (the SHELL 1 successor), using the
    6-transcendental floor proven small + phi_fold (proven running)
