diff --git a/AGENTS.md b/AGENTS.md
index 1e11b77..144596b 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -25,6 +25,7 @@ The root `Makefile` is a lean delegator: base commands (`run`, `build`, `test`,
 - Odin code lives in `gui/src/`; external deps go in `gui/lib/` and are wired via `-collection:lib=lib/local` (or a git submodule imported by relative path).
 - Everything is plain `make` — no Taskfile — so CI (Gitea Actions) can call `make` directly.
 - Keep the renderer (gui/), API (api/), and infra (infra/) logically separated; each owns its own Makefile, and the root Makefile is the only place that ties them together.
+- **Diagrams in this repo's documentation are Mermaid flowcharts.** Gitea renders ` ```mermaid ` fenced blocks natively. Prefer a Mermaid flowchart over ASCII art / box-drawing diagrams; if a diagram genuinely can't be expressed as a flowchart, fall back to a plain text-based markdown diagram (e.g. a code block or table) rather than hand-rawn ASCII boxes.
 
 ## Gotchas
 - Odin version is pinned in `.gitea/workflows/*.yml` (`ODIN_VERSION`) and defaults in `scripts/install_odin.sh`; bump both together when tracking a new release.
diff --git a/resources/ai/research/data-streaming/01-architecture.md b/resources/ai/research/data-streaming/01-architecture.md
index b649458..3504594 100644
--- a/resources/ai/research/data-streaming/01-architecture.md
+++ b/resources/ai/research/data-streaming/01-architecture.md
@@ -59,11 +59,13 @@ get_catalog_objects :: proc(url: string, catalog_name: string)            // nil
 
 ## Data flow gap
 
-```
-[DESI catalog store]  --(future)-->  [Rust/axum API]  --(nothing today)-->  [Odin + raylib GUI]
-                                        ^                                        ^
-                                        | serde JSON models                    | hand-mirrored structs
-                                        |                                      | (stubs, never used)
+```mermaid
+flowchart LR
+    A["DESI catalog store"]
+    B["Rust / axum API<br/><i>serde JSON models</i>"]
+    C["Odin + raylib GUI<br/><i>hand-mirrored structs<br/>stubs, never used</i>"]
+    A -. "future ingestion" .-> B
+    B --x|"nothing today"| C
 ```
 
 There is **no live data flow**. The API currently returns JSON placeholders; the
@@ -116,17 +118,16 @@ response protocol with a cheap binary payload would fit this well.
 
 ## High-level target architecture
 
-```
-        catalog.fbs (single source of truth, checked into repo)
-              |
-     +--------+---------+
-     |                  |
-flatc --rust       flatcc --c (or hand-rolled Odin reader)
-     |                  |
-api/ (Rust)        gui/ (Odin + raylib)
-     |                  ^
-     |   HTTP / WebSocket (framed FlatBuffer binary stream)
-     +------------------+
+```mermaid
+flowchart TB
+    S["catalog.fbs<br/><i>single source of truth<br/>checked into repo</i>"]
+    R["flatc --rust"]
+    C["flatcc --c<br/><i>or hand-rolled Odin reader</i>"]
+    API["api/ · Rust"]
+    GUI["gui/ · Odin + raylib"]
+    S --> R --> API
+    S --> C --> GUI
+    API <-->|"HTTP / WebSocket<br/>framed FlatBuffer binary stream"| GUI
 ```
 
 - One schema file. Two generators. Byte-for-byte identical wire format.
diff --git a/resources/ai/research/data-streaming/02-flatbuffers-overview.md b/resources/ai/research/data-streaming/02-flatbuffers-overview.md
index e3549c3..916b423 100644
--- a/resources/ai/research/data-streaming/02-flatbuffers-overview.md
+++ b/resources/ai/research/data-streaming/02-flatbuffers-overview.md
@@ -105,19 +105,28 @@ These are the "thou shalt" rules for keeping buffers compatible:
 
 ## Reading a buffer (conceptual)
 
+Buffer layout:
+
+```mermaid
+flowchart LR
+    subgraph BUF["bytes: &[u8]"]
+        O["uoffset<br/>root table offset"]
+        FI["file_identifier<br/>(optional)"]
+        D["tables · vtables · data"]
+    end
+    O --> FI --> D
 ```
-bytes: &[u8]
 
-  ┌─────────────────────────────┐
-  │ uoffset (root table offset)  │
-  │ file_identifier (optional)   │
-  │ ... tables, vtables, data ...│
-  └─────────────────────────────┘
+Access sequence — each field is a few offset dereferences and a read:
 
-root = follow(bytes)              // jump to root table via uoffset
-vtable = root - root.vtable_off  // locate vtable for this table
-field_ra  = vtable.slot_ra != 0  // present?
-if present: ra = read_f64(bytes, root + slot_ra)
+```mermaid
+flowchart TD
+    R["root = follow(bytes)<br/>jump to root table via uoffset"]
+    V["vtable = root − root.vtable_off<br/>locate vtable for this table"]
+    Q{"field slot present?"}
+    R --> V --> Q
+    Q -- "no" --> DEF["use schema default"]
+    Q -- "yes" --> RD["ra = read_f64(bytes, root + slot_ra)"]
 ```
 
 There is **no parsing loop**. Each accessor is a few offset dereferences and a
diff --git a/resources/ai/research/data-streaming/03-rust-integration.md b/resources/ai/research/data-streaming/03-rust-integration.md
index 4629d58..acc0b9d 100644
--- a/resources/ai/research/data-streaming/03-rust-integration.md
+++ b/resources/ai/research/data-streaming/03-rust-integration.md
@@ -196,6 +196,18 @@ Notes:
 
 ## Building buffers efficiently (Rust specifics)
 
+Build order is **back-to-front** (children before parents):
+
+```mermaid
+flowchart TB
+    A["create_string / create_vector<br/>children first"]
+    B["create nested child tables"]
+    C["create parent table<br/>ObjectBatch::create(&args)"]
+    D["builder.finish(root, Some(&#34;DESI&#34;))"]
+    E["Bytes::copy_from_slice(fbb.finished_data())<br/>→ HTTP / WebSocket response"]
+    A --> B --> C --> D --> E
+```
+
 - `FlatBufferBuilder::with_capacity(n)` pre-allocates; `reset()` reuses the
   buffer across messages. In a loop streaming batches, create one builder, reuse
   it — avoid repeated reallocation.
diff --git a/resources/ai/research/data-streaming/04-odin-client-integration.md b/resources/ai/research/data-streaming/04-odin-client-integration.md
index 52f9c98..c406bd6 100644
--- a/resources/ai/research/data-streaming/04-odin-client-integration.md
+++ b/resources/ai/research/data-streaming/04-odin-client-integration.md
@@ -17,6 +17,21 @@ has two C-related access points:
    schema plus a small `libflatccrt.a` runtime. Works via the C ABI, so Odin's
    `foreign import` can consume it.
 
+Choosing a path:
+
+```mermaid
+flowchart TD
+    NAT{"primary target is native desktop?"}
+    NAT -- "yes" --> CGO{"want to avoid C in the build?"}
+    CGO -- "yes" --> PATHA["Path B · pure-Odin reader<br/>hand-rolled, no C dependency"]
+    CGO -- "no" --> PATHA
+    CGO -- "prefer proven lib / less maintenance" --> PATHC["Path A · FFI to FlatCC<br/>bind generated C headers"]
+    PATHC --> REUSE["Path C · OdinArrow reuse<br/>or borrow its decode patterns"]
+    NAT -- "no · browser/WASM" --> PATHD["Path D · TS/JS interop<br/>official JS lib → typed arrays into WASM"]
+```
+
+Index of paths:
+
 | Path | Effort | Zero-copy on reads | Notes |
 |---|---|---|---|
 | A: FFI to FlatCC (C runtime) | Medium | ✅ | Bind generated C headers to Odin `foreign` |
@@ -175,12 +190,14 @@ typed Odin slices.
 The rendering win only materializes if data stays zero-copy **into the frame
 loop**:
 
-1. Fetch frame bytes → owned `[dynamic]u8` (or a slice pinned for the lifetime
-   of the frame).
-2. `verify` the buffer once.
-3. Get `ra_slice := ObjectBatch.ra(&buf)` → `[]f64` view.
-4. Per object in `update()`/`draw()`: read `ra[i]`, `dec[i]`, `z[i]` straight
-   from that slice; build `rl.Vector3`; `DrawPoint3D`.
+```mermaid
+flowchart LR
+    A["fetch frame bytes → [dynamic]u8<br/>or a slice pinned for the frame"]
+    B["verify the buffer once"]
+    C["ObjectBatch.ra(&buf) → []f64 view"]
+    D["per object in update()/draw()<br/>ra[i] · dec[i] · z[i] → rl.Vector3 → DrawPoint3D"]
+    A --> B --> C --> D
+```
 
 No per-object allocation. The current `Galaxy { position, color }` dynamic array
 in `main.odin` is the data structure you'd replace with *slices into the
diff --git a/resources/ai/research/data-streaming/05-streaming-protocols.md b/resources/ai/research/data-streaming/05-streaming-protocols.md
index 89a5798..76b8f48 100644
--- a/resources/ai/research/data-streaming/05-streaming-protocols.md
+++ b/resources/ai/research/data-streaming/05-streaming-protocols.md
@@ -24,14 +24,17 @@ offers two built-in mechanisms plus the community pattern:
 
 ### Option 1: Size-prefixed FlatBuffers (built-in)
 
+```mermaid
+flowchart LR
+    A["u32 LE<br/>total buffer len<br/><i>size prefix</i>"]
+    B["u32 LE<br/>root table offset"]
+    C["file identifier<br/>(4 bytes)"]
+    D["tables · vtables · data"]
+    A --> B --> C --> D
+```
+
 ```rust
 builder.finish_size_prefixed(root, Some("DESI"));
-// +---------------------------+
-// | u32 LE: total buffer len  |  <-- size prefix
-// | u32 LE: root table offset |
-// | file identifier (4 bytes) |
-// | ... data ...              |
-// +---------------------------+
 ```
 
 Reader side:
@@ -53,8 +56,12 @@ message kinds).
 
 ### Option 2: Custom length-prefix framing (like `flatstream`)
 
-```
-[ u32 LE: message_len ] [ optional checksum (e.g. u32 crc/xxhash) ] [ flatbuffer payload ]
+```mermaid
+flowchart LR
+    A["u32 LE<br/>message_len"]
+    B["optional checksum<br/>(u32 crc / xxhash)"]
+    C["FlatBuffer payload"]
+    A --> B --> C
 ```
 
 - `flatstream-rs` (see `03-rust-integration.md`) is a reference implementation
@@ -260,6 +267,16 @@ FlatBuffers long-term (mmap-friendly, page-in-what-you-touch).
 
 ## Decision summary for this repo
 
+```mermaid
+flowchart TD
+    A["HTTP GET → one FlatBuffer body per batch<br/>validate Rust builder + Odin reader"]
+    B["WebSocket → one Binary message per batch<br/>interactive path · no custom framing"]
+    C["Self-identifying messages<br/>file_identifier &#34;DESI&#34;"]
+    D["size-prefixed / flatstream-style framing<br/>or HTTP-range + mmap for static catalogs"]
+    A --> B --> C
+    C -. "later, if needed" .-> D
+```
+
 1. Start with **HTTP GET → one FlatBuffer body per batch** to validate the Rust
    builder + Odin reader (no protocol work at all).
 2. Then add **WebSocket** with one `Binary` message per batch (no custom framing)
diff --git a/resources/ai/research/data-streaming/07-alternatives.md b/resources/ai/research/data-streaming/07-alternatives.md
index cf16a6e..d6761ff 100644
--- a/resources/ai/research/data-streaming/07-alternatives.md
+++ b/resources/ai/research/data-streaming/07-alternatives.md
@@ -108,6 +108,15 @@ message shape — see `10-performance-benchmarks.md`.)
 
 ## Bottom line
 
+```mermaid
+flowchart TD
+    Q1{"zero-copy reads<br/>in the per-frame render loop?"}
+    Q1 -- "no" --> PB["Protobuf / gRPC<br/>decode once into draw buffers"]
+    Q1 -- "yes" --> Q2{"truly columnar?<br/>millions of rows"}
+    Q2 -- "yes" --> ARR["Apache Arrow IPC<br/>via OdinArrow"]
+    Q2 -- "no · batched vectors" --> FB["FlatBuffers · this proposal"]
+```
+
 - **FlatBuffers is the best default** for this project: the zero-copy read model
   matches the render loop, the wire format is compact for numeric vectors, schema
   evolution fits DESI's release cadence, and the Rust + WASM/JS official story
diff --git a/resources/ai/research/data-streaming/08-testing-strategies.md b/resources/ai/research/data-streaming/08-testing-strategies.md
index 2aaf26b..67bb3c0 100644
--- a/resources/ai/research/data-streaming/08-testing-strategies.md
+++ b/resources/ai/research/data-streaming/08-testing-strategies.md
@@ -139,7 +139,23 @@ side / flatc* — see the "cross-language fixture" section below.
 
 ### 4. Cross-language conformance suite (THE key integration test)
 
-This is the test that actually catches incompatibility. Design:
+This is the test that actually catches incompatibility. Pipeline:
+
+```mermaid
+flowchart LR
+    S["schema/catalog.fbs"]
+    J["testdata/catalog_sample.json"]
+    S --> F["flatc --binary"]
+    J --> F
+    F --> BIN["committed .bin fixtures<br/>repo-checked-in"]
+    BIN --> OT["Odin tests<br/>assert identical values"]
+    BIN --> RT["Rust tests<br/>assert expected values"]
+    RT -. "deterministic builder" .-> PAR["byte-for-byte parity"]
+    OT -. "reads it" .-> PAR
+    PAR -. "catch drift" .-> F
+```
+
+Design:
 
 1. **Static fixtures, committed to the repo** (`testdata/*.bin`):
    - Built once by `flatc --binary <schema>.fbs <data>.json` (deterministic,
diff --git a/resources/ai/research/data-streaming/README.md b/resources/ai/research/data-streaming/README.md
index 79848d7..f308b34 100644
--- a/resources/ai/research/data-streaming/README.md
+++ b/resources/ai/research/data-streaming/README.md
@@ -70,6 +70,16 @@ directly from the buffer each frame without allocations.
 
 ## Immediate Next Steps (when you're ready to implement)
 
+```mermaid
+flowchart TD
+    A["Prototype schema<br/>catalog.fbs: Catalog · CatalogObject · ServerMessage union"]
+    B["Generate Rust code<br/>flatc --rust → api/build.rs · serve WS via axum"]
+    C["Prototype the Odin reader<br/>flatcc FFI · pure-Odin · OdinArrow"]
+    D["Static fixture files<br/>flatc --binary → committed .bin"]
+    E["Cross-language tests<br/>Rust + Odin read the same fixtures identically"]
+    A --> B --> C --> D --> E
+```
+
 1. **Prototype schema first.** Write `catalog.fbs` covering `Catalog`, `CatalogObject`,
    and a `ServerMessage` union (handshake / catalog list / chunk of objects / end).
 2. **Generate Rust code** via `flatc --rust` in an `api/build.rs` (see