diff --git a/docs/rust-port.md b/docs/rust-port.md
new file mode 100644
index 0000000..c2784f5
--- /dev/null
+++ b/docs/rust-port.md
@@ -0,0 +1,70 @@
+# Rust port — staged plan
+
+This is the plan for the native-Rust direction first discussed alongside the A/B-bootloader
+idea. **What changed since then:** the track format is now formally specced
+(`docs/track-format.md`) with a golden-vector conformance suite (`tests/run.mjs`). That suite is
+the thing that makes a port *safe* — any Rust engine has a precise, executable definition of
+"correct" to validate against, the same one `engine.js` and `app.py` already pass.
+
+## The core idea
+
+Port from the **inside out**, lowest-risk first. The pure logic (track codec, then the
+scheduler) is host-testable with zero hardware and is gated by the existing golden vectors. Only
+once that's proven do we touch drivers, A/B, and the actual firmware. We do **not** rip out
+CircuitPython until the Rust engine passes the vectors *and* the drivers are proven on hardware.
+
+## Stages
+
+### Stage 0 — toolchain in a container
+Add a Rust toolchain image (mirroring `hardware/eda/`): a `Containerfile` with `rustup`, the
+`thumbv8m.main-none-eabihf` target (RP2350 is Cortex-M33), `flip-link`, `probe-rs`, `elf2uf2`.
+Driven by a `run.sh` like the EDA one. **Never on the host.**
+
+### Stage 1 — `track-format` crate  ← the concrete first PR
+A pure, `no_std`-compatible crate: `parse(&str) -> Track` and `serialize(&Track) -> String`,
+plus a `normalize()` that emits the neutral structure from `docs/track-format.md` §5. Then a
+`cargo test` that reads `tests/fixtures/track-format.json` and asserts each case's `norm` and
+round-trip — i.e. a **third adapter alongside `js_adapter.mjs` / `py_adapter.py`**. When this is
+green, the Rust engine provably agrees with web + device on every groove, euclid, swing, ghost,
+polymeter, and the playback-flow tokens. No hardware, fully testable in the container.
+
+This is the highest-value slice: small, gated by work already done, and it proves the toolchain.
+
+### Stage 2 — scheduler/engine
+Port the look-ahead step scheduler. `engine.js` already marks these primitives
+`PORTS TO FIRMWARE`, and `app.py`'s `tick()` is the same model. Host-test it by asserting click
+*timings* for known patches (e.g. swing ratios, polymeter bar lengths) — still no hardware.
+
+### Stage 3 — drivers (hardware)
+On `embassy` / `rp-hal`:
+- ST7789 240×320 display → `mipidsi` + `embedded-graphics` (mature; the parts are well-supported).
+- I²S to the PCM5102A → RP2350 PIO.
+- WS2812 → `ws2812-pio`.  USB-MIDI → `usbd-midi` / `embassy-usb`.
+- GT911 touch (Kit) over I²C.
+
+### Stage 4 — native A/B + secure boot
+Replace the `.mpy`-level A/B hack (`code.py` loads `app.mpy`, rolls back to `app.bak`) with the
+**RP2350 bootrom's native** partition-table A/B + signed boot, configured via `picotool` (the
+chip already provides this — see the earlier hardware discussion). The Rust app is the image in
+the slot; rollback and version selection move into silicon.
+
+## What you keep / lose
+
+- **Gain:** memory safety, native A/B + secure boot, performance headroom, one typed model instead
+  of three hand-written parsers.
+- **Lose:** the live one-click `.mpy` push (Rust is compile→flash→reboot). The editor's *data*
+  live-sync (tempo/pattern/setlist mirroring) still works — it's a data protocol. Only live
+  *logic* edits go away, and an embedded `wasm3`/script module could buy those back if wanted.
+
+## Acceptance gate
+
+Every codec/engine change must pass `tests/fixtures/track-format.json`. The Rust crate joins
+`js`/`py` as a runner adapter, so "same groove on web, device, and the Rust build" is enforced,
+not hoped for.
+
+## Recommendation
+
+Do **Stage 1 in a container next** — it's small, testable today (given a toolchain), reuses the
+suite, and produces a real artifact to judge the Rust path on before committing to drivers or a
+firmware rewrite. Defer Stages 3–4 until Stage 1–2 are green and you've decided the live-push
+tradeoff is acceptable.