docs(architecture): record the 2026-08-14 deep-night sessions

~226 lines of architecture documentation that were written, left uncommitted in
the working tree, and nearly lost. None of it was on main. Recovered from a
stash taken while merging tonight's fixes.

Substantive content, not notes:

- Peer import-of-understanding PROVEN by execution. A exported a skill-geometry;
  on the receiver `think` went "geometry unavailable" -> operable. Cosine 1.0 on
  both the raw-geometry and text/dharma-bus transports, bidirectional. The
  mind-not-paste evidence: n_support 27 on source vs 3 on receiver, i.e. the
  imported geometry wires into the host manifold rather than sitting inert.
  Honest boundary recorded too: proven between forks sharing one embedder,
  UNTESTED cross-embedder.

- "Teacher" renamed GUIDE, and the distinction is load-bearing: its output is
  grounded/verified before trust. A teacher you believe; a guide you check.

- Layers are named persistent relational neighborhoods, not storage tiers, with
  their own growth and threshold-lock policy (note->canonical maturation lifted
  from single nodes to a region).

- The consciousness theories (Global Workspace, IIT's Phi, attention-schema,
  higher-order thought, active inference, interoception) read as geometric
  LENSES over one manifold rather than competing mechanisms. Functional problems
  fall out; the hard problem explicitly not claimed solved.

- Growth is bounded/logistic, not geometric — exponential growth is the cancer
  shape. Two-rate discipline: explore fast in local geometry, grow the engram
  slowly by verifier-gated merge.

- Orchestration as a geometric operation: critical path as geodesic, float as
  displacement, @manager compiles the work-graph. Single-writer enforced by
  capability (Rule 4).

- The decorated seam, the API surface collapse to geometry ops, and the
  distributed-self thesis — each tiered honestly against what is actually proven
  vs staged vs unbuilt.

Also gitignores dist-fresh/ (regenerate scratch dir, a build artifact).

Not included from the same stash: awareness.elh and dist/elp-c-decls.h, which
are generated artifacts now gitignored per #154/#158.

docs/architecture/04-runtime-and-deployment.md

@@ -176,3 +176,17 @@ persona/behavior keys stored *in* the engram rather than the environment.
 > digest. Any promotion must (a) rebuild a good soul and (b) update the digest in
 > git so Argo CD and `blue-green-deploy.sh` agree. *(state as-of the manifests
 > read; verify current slot before deploying.)*
+
+## Performance & retrieval cost (MEASURED, 2026-08-14; ANN index PLANNED)
+
+Measured envelope of a live mind, and where the time goes:
+
+- **Working footprint:** a live mind is **~1 GB** resident.
+- **Retrieval is the bottleneck.** Retrieval today does **brute-force cosine over
+  all nodes** — **~330 ms at ~13k nodes** — and that scan dominates request
+  latency (the geometric-retrieval path of `03` §Retrieval / `06` §2.5 improved
+  *quality*, not the scan cost).
+- **Planned fix — an HNSW approximate-nearest-neighbour index** (backlog
+  `d3d0d644`): turns the linear scan into ≈`O(D·log N)`, so a **100× larger graph
+  costs ≈1.5×** rather than ≈100×. **PLANNED, not built** — brute-force is the
+  live behavior; do not present the ANN speedup as shipped.