diff --git a/docs/track-format.md b/docs/track-format.md
new file mode 100644
index 0000000..ee6d15c
--- /dev/null
+++ b/docs/track-format.md
@@ -0,0 +1,207 @@
+# PM track format — specification
+
+This is the **single source of truth** for the metronome's track ("program" / "patch")
+format. The grammar is currently implemented by hand in three places that have quietly
+drifted:
+
+- **Web** — `src/engine.js` (`patchToSetup` / `laneStrToCfg` / `setupToPatch` / `laneCfgToStr`)
+- **Firmware** — `pico-cp/app.py` (`parse_program` / `_parse_lane` / `lane_to_str` / `_prog_str`)
+- (`editor.html` inlines `engine.js` at build time)
+
+`tests/fixtures/track-format.json` holds golden vectors that pin every feature to a single
+expected meaning. `tests/run.mjs` runs both the JS and Python implementations against them
+and reports divergences. **Any new implementation (e.g. a Rust engine) must pass the same
+vectors** — that is what keeps "the same groove on the device and in the browser" true.
+
+Status legend used below and in the fixtures:
+
+- **stable** — implemented and (intended to be) identical on web + firmware.
+- **divergence** — a real cross-implementation disagreement that exists *today*; the vector
+  encodes the spec's intended behavior and the runner flags the side that is wrong.
+- **new** — defined here but not yet implemented anywhere (the playback-flow model below);
+  the vector is the acceptance test for building it.
+
+---
+
+## 1. Container: `programs.json`
+
+The device reads set-lists from `/programs.json` (pushed by the editor over USB-MIDI SysEx,
+or dragged onto the CIRCUITPY drive). Built-in set-lists are baked into firmware and are *not*
+in this file.
+
+```jsonc
+{
+  "format": 2,                       // NEW. absent ⇒ treat as format 1 (legacy)
+  "setlists": [
+    {
+      "title": "My set",
+      "description": "optional",
+      "onEnd": "stop",               // NEW: stop | nextList | loop   (default: stop)
+      "defaultEnd": "next",          // NEW: items without their own end= inherit this
+      "programs": [
+        { "name": "Intro", "prog": "t88;b8;kick:4=X.x.;end=next" },
+        { "name": "Groove", "prog": "t88;kick:4;snare:4=.X.X" }
+      ]
+    }
+  ]
+}
+```
+
+**Legacy (format 1):** `{ "setlists":[...] }` with no `format`, or the older flat
+`{ "programs":[...] }` (a single list). Loaders MUST still accept both. Migration to format 2:
+set `format: 2`; a previously "continuous" set-list gets `onEnd` per the author's intent and
+`defaultEnd: "next"` (or `end=next` on each item).
+
+---
+
+## 2. Patch string grammar
+
+A patch is `;`-separated tokens. **Rule that prevents token collisions:** a token containing
+`:` is a **lane**; every other token is a **keyword directive** matched against the reserved
+set `{ v1, t, vol, cd, b, tr, rmp, rep, end }`. Unknown tokens are ignored (see §6).
+
+```ebnf
+patch      = [ "v1" ";" ] directive *( ";" directive ) ;
+
+directive  = tempo | volume | countin | bars | trainer | ramp | rep | end | lane ;
+
+tempo      = "t" int ;                       (* beats per minute, clamped to 5..300 *)
+volume     = "vol" int ;                      (* 0..100, host-optional — device ignores *)
+countin    = "cd" int ;                       (* count-in seconds, host-optional *)
+bars       = "b" int ;                         (* cycle length in bars; drives Continue/rep *)
+trainer    = "tr" int "/" int ;                (* playBars "/" muteBars (gap trainer) *)
+ramp       = "rmp" int "/" signed "/" int ;    (* startBpm "/" amount "/" everyBars *)
+
+rep        = "rep" int ;                        (* NEW: cycles before end fires; default 1 *)
+end        = "end" "=" ( "stop" | "next" | signed ) ;  (* NEW: see §3 *)
+
+lane       = sound ":" groups [ "/" sub [ "s" ] ] [ euclid ] [ "=" pattern ]
+                    [ "@" signed ] [ "~" ] [ "!" ] ;
+sound      = name | int ;                       (* int = GM percussion note number alias *)
+groups     = int *( "+" int ) ;                 (* "4"  or  "2+2+3" → beats per bar *)
+sub        = int ;                              (* subdivision; trailing "s" = swing *)
+euclid     = "(" int [ "," int [ "," int ] ] ")" ;   (* k [, n [, rot ]] — even distribution *)
+pattern    = *( "X" | "x" | "g" | "." | "-" | "_" ) ; (* per-step dynamics *)
+```
+
+### Lane semantics
+
+- **groups** — `2+2+3` means a 7-beat bar grouped 2+2+3; the first step of each group is
+  accented by default. `beatsPerBar = sum(groups)`.
+- **sub** — steps per beat (`/2` eighths, `/3` triplets, `/4` sixteenths). `steps = beats*sub`.
+- **swing** — `/2s` lays the off-beat on the last triplet (≈ 2/3).
+- **pattern** — one char per step: `X`=accent (level 2), `x`/`1`=normal (1), `g`=ghost (3),
+  `.`/`-`/`_`/anything else = rest (0). Short patterns are right-padded with rests to `steps`.
+  With no pattern, the first step of each beat is accented and the rest are normal.
+- **euclid** `(k,n,rot)** — `k` hits spread as evenly as possible over `n` steps (rotated by
+  `rot`), first hit accented. Replaces an explicit `=pattern`.
+- **`@<db>`** — per-lane gain in dB. (Host-optional: applied in the browser; see §6.)
+- **`~`** — polymeter: this lane keeps its own bar length independent of the master lane.
+- **`!`** — lane present but silenced/disabled.
+
+### Top-level directives
+
+- **`t`** tempo, clamped to `[5, 300]`.
+- **`b<n>`** the segment's **cycle length** in bars — used for Continue/`rep` accounting and
+  song auto-advance. Absent ⇒ the cycle is one bar of the **master lane** (the first lane).
+- **`rmp<start>/<amt>/<every>`** tempo ramp: every `<every>` bars, change tempo by `<amt>`
+  (may be negative), starting from `<start>`.
+- **`tr<play>/<mute>`** gap trainer: play `<play>` bars, silence `<mute>` bars, repeat.
+- **`vol`, `cd`** host-optional (the device has its own volume and no count-in).
+
+---
+
+## 3. Playback flow (NEW)
+
+Replaces the old **global** `Continue` toggle with **per-track** behavior.
+
+**Default (no `end=` token) — loop forever**, exactly like a metronome. Manual advance
+(joystick / footswitch) always moves to the next track. "Vamp until cue" is therefore the
+default and needs no special token.
+
+`end=` exists only to make flow **automatic** after a finite number of cycles:
+
+| Token            | Meaning                                                              |
+|------------------|---------------------------------------------------------------------|
+| *(absent)*       | Loop forever. Manual advance → next track.                          |
+| `end=stop`       | After `rep` cycles, stop.                                            |
+| `end=next`       | After `rep` cycles, auto-advance one track. (Sugar for `end=+1`.)    |
+| `end=<±N>`       | **Relative goto**: after `rep` cycles, jump `N` tracks (e.g. `end=-2` repeats a section, D.S.). |
+| `rep=<N>`        | Cycles to play before `end` fires. Default `1`. Only meaningful with `end`. |
+
+A **cycle** = `b<bars>` bars if `b` is present, else one bar of the master lane.
+
+**Normalization** (what the golden vectors compare): `end` normalizes to `"stop"`, or an
+integer offset (`next` ⇒ `1`); absent ⇒ `null`. `rep` ⇒ the integer (defaults to `1` when
+`end` is present), else `null`.
+
+### Resolution & boundaries
+
+- **Manual override always wins.** `end=` is only the hands-off behavior.
+- **Seam is gapless.** An automatic `next`/goto does **not** re-trigger count-in and inserts
+  no gap (honors the existing `_seam_t` seam). Count-in (`cd`) fires only on manual/initial
+  start. Tempo jumps to the destination track's `t` at the seam.
+- **Goto past the last item** (or `end=next` on the last item) triggers the set-list's
+  `onEnd` policy: `stop` | `nextList` | `loop`.
+- **Goto before the first item** clamps to the first item.
+- **`defaultEnd`** on a set-list is inherited by any item whose patch has no `end=`.
+
+---
+
+## 4. Canonical form & round-tripping
+
+Serializing a parsed patch MUST be **idempotent**: `serialize(parse(serialize(parse(x))))`
+equals `serialize(parse(x))`. Host-optional fields that a host does not act on
+(`vol`, `cd`, `@db`) MUST still survive the round-trip rather than being dropped — silent
+data loss is the failure mode this spec exists to prevent.
+
+The runner checks **semantic equality** (the normalized structure — what actually plays)
+across implementations, plus per-implementation idempotency. It does **not** require the two
+serializers to emit byte-identical strings (they legitimately differ on the optional `v1`
+prefix and on whether a default pattern is written out).
+
+---
+
+## 5. Normalized structure (the comparison target)
+
+Each implementation's adapter parses a patch and emits this neutral shape; vectors store the
+expected value in `norm`:
+
+```jsonc
+{
+  "bpm": 120, "bars": 0, "volume": null, "countMs": 0,
+  "ramp":    null,            // or { "start": 80, "amt": 4, "every": 4 }
+  "trainer": null,            // or { "play": 2, "mute": 2 }
+  "rep":     null,            // NEW
+  "end":     null,            // NEW: "stop" | <int> | null
+  "lanes": [
+    { "sound": "kick", "groups": [4], "sub": 1, "swing": false,
+      "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] }
+  ]
+}
+```
+
+`levels` is the resolved per-step dynamics array (0 rest / 1 normal / 2 accent / 3 ghost) —
+the real audible payload, and the most important thing two implementations must agree on.
+
+---
+
+## 6. Known divergences (today) — the work list
+
+Surfaced by the runner; each has a `divergence` vector:
+
+1. **Euclid `(k,n,rot)`** — parsed by `engine.js`, **not** by firmware (`_parse_lane` can't
+   strip `(...)`, so it falls back to a plain unaccented bar). → Implement on device.
+2. **`vol` / `cd`** — round-tripped by `engine.js`, **dropped** by firmware. → Firmware should
+   preserve them on serialize even though it ignores them at runtime.
+3. **Unknown sound name** — `engine.js` falls back to `beep`; firmware keeps the raw name.
+   → Spec: unknown sound ⇒ `beep` everywhere.
+4. **Tempo clamp** — firmware clamps `t` to `[5,300]`; `engine.js` does not. → Clamp everywhere.
+5. **Empty patch** — firmware injects a default `beep:4` lane; `engine.js` returns zero lanes
+   (the editor guards emptiness separately). → Spec: ≥1 lane; empty ⇒ `beep:4`.
+6. **`@db` gain** — parsed on both sides but **applied** only in the browser (runtime, not a
+   parse divergence; not a fixture failure — listed for completeness).
+
+The capability/version handshake (the device already replies with its firmware version over
+SysEx) should gate features so the editor can warn when a track uses something the connected
+firmware is too old to play, instead of letting it degrade silently.
diff --git a/tests/.gitignore b/tests/.gitignore
new file mode 100644
index 0000000..7a60b85
--- /dev/null
+++ b/tests/.gitignore
@@ -0,0 +1,2 @@
+__pycache__/
+*.pyc
diff --git a/tests/README.md b/tests/README.md
new file mode 100644
index 0000000..1d264e6
--- /dev/null
+++ b/tests/README.md
@@ -0,0 +1,40 @@
+# Track-format conformance tests
+
+Golden-vector suite that pins the track ("program"/"patch") format to a single meaning and
+checks that both implementations agree:
+
+- **web** — `src/engine.js`
+- **firmware** — `pico-cp/app.py`
+
+The spec is `docs/track-format.md`. Any new implementation (e.g. a Rust engine) must pass the
+same vectors — that is what keeps "the same groove on the device and in the browser" true.
+
+## Run
+
+```sh
+node tests/run.mjs        # table of pass / known-divergence / FAIL per case
+node tests/run.mjs -v     # also print expected-vs-actual diffs for unexpected failures
+```
+
+Exit code is non-zero on any **unexpected** failure or round-trip (idempotency) break, so it
+works as a CI gate.
+
+## Layout
+
+- `fixtures/track-format.json` — the vectors. Each has `in` (a patch), `norm` (expected
+  normalized meaning, see spec §5), a `status`, and optional `expectFail` listing impls known
+  to differ today.
+- `adapters/js_adapter.mjs` — loads the real `src/engine.js` grammar (no copy) and normalizes.
+- `adapters/py_adapter.py` — extracts the real `pico-cp/app.py` grammar functions via `ast`
+  (no copy) and normalizes.
+- `run.mjs` — runs every vector through both adapters and reports.
+
+## Reading the result
+
+- `✓ pass` — implementation matches the spec for that vector.
+- `· known` — a divergence/feature listed in `expectFail`; expected, not a failure.
+- `✗ FAIL` — an **unexpected** mismatch (a regression). Investigate.
+- `★ fixed` — an impl listed in `expectFail` now passes; remove it from `expectFail`.
+
+When you fix a divergence in code, delete that impl from the case's `expectFail`. When you
+implement the `new` playback-flow tokens (`rep` / `end`), those cases flip to `pass`.
diff --git a/tests/adapters/js_adapter.mjs b/tests/adapters/js_adapter.mjs
new file mode 100644
index 0000000..2c70e67
--- /dev/null
+++ b/tests/adapters/js_adapter.mjs
@@ -0,0 +1,51 @@
+// JS adapter: loads the REAL grammar out of src/engine.js (no copy) and emits the
+// neutral normalized structure defined in docs/track-format.md §5.
+//
+// engine.js has no top-level side effects (only const/function declarations) and references
+// `window` only inside ensureAudio(), which we never call — so it loads cleanly in Node.
+import { readFileSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+import { dirname, join } from "node:path";
+
+const here = dirname(fileURLToPath(import.meta.url));
+const src = readFileSync(join(here, "..", "..", "src", "engine.js"), "utf8");
+
+// Expose the pure codec functions from engine.js's top-level scope.
+globalThis.window = globalThis.window || {};
+const api = new Function(src + "\n;return { patchToSetup, setupToPatch };")();
+
+const groupsArr = (s) => String(s).split(/[^0-9]+/).filter(Boolean).map(Number);
+
+export function normalize(patch) {
+  const s = api.patchToSetup(patch);
+  return {
+    bpm: s.bpm,
+    bars: s.bars || 0,
+    volume: s.volume == null ? null : s.volume,
+    countMs: s.countMs || 0,
+    ramp: s.ramp && s.ramp.on ? { start: s.ramp.startBpm, amt: s.ramp.amount, every: s.ramp.everyBars } : null,
+    trainer: s.trainer && s.trainer.on ? { play: s.trainer.playBars, mute: s.trainer.muteBars } : null,
+    rep: s.rep == null ? null : s.rep,   // not parsed yet → undefined → null
+    end: s.end == null ? null : s.end,   // not parsed yet → undefined → null
+    lanes: (s.lanes || []).map((c) => ({
+      sound: c.sound,
+      groups: groupsArr(c.groupsStr),
+      sub: c.stepsPerBeat || 1,
+      swing: !!c.swing,
+      poly: !!c.poly,
+      mute: c.enabled === false,
+      gainDb: c.gainDb || 0,
+      levels: (c.beatsOn || []).map((v) => v | 0),
+    })),
+  };
+}
+
+// serialize(parse(x)) — used for the idempotency check.
+export function canonical(patch) {
+  return api.setupToPatch(api.patchToSetup(patch));
+}
+
+// CLI: `node js_adapter.mjs '<patch>'` prints normalized JSON.
+if (process.argv[1] && fileURLToPath(import.meta.url) === process.argv[1]) {
+  process.stdout.write(JSON.stringify(normalize(process.argv[2] ?? "")));
+}
diff --git a/tests/adapters/py_adapter.py b/tests/adapters/py_adapter.py
new file mode 100644
index 0000000..3ef8989
--- /dev/null
+++ b/tests/adapters/py_adapter.py
@@ -0,0 +1,93 @@
+#!/usr/bin/env python3
+"""Python adapter: extracts the REAL grammar functions out of pico-cp/app.py (no copy) and
+emits the neutral normalized structure from docs/track-format.md §5.
+
+app.py can't be imported directly (it does hardware init at module load), so we pull just the
+pure codec nodes by name via `ast` and exec them in isolation. This tests the actual firmware
+code and survives edits as long as the function/constant names persist.
+"""
+import ast
+import json
+import os
+import sys
+
+APP = os.path.join(os.path.dirname(__file__), "..", "..", "pico-cp", "app.py")
+WANT = {"PAT", "PRIO", "PAT_CH", "parse_program", "_parse_lane", "lane_to_str"}
+
+with open(APP) as f:
+    _src = f.read()
+
+_segs = []
+for node in ast.parse(_src).body:
+    name = None
+    if isinstance(node, ast.Assign) and len(node.targets) == 1 and isinstance(node.targets[0], ast.Name):
+        name = node.targets[0].id
+    elif isinstance(node, ast.FunctionDef):
+        name = node.name
+    if name in WANT:
+        _segs.append(ast.get_source_segment(_src, node))
+
+NS = {}
+exec("\n".join(_segs), NS)
+
+
+def _prog_str(bpm, lanes, bars, ramp, trainer):
+    # Mirrors app.py App._prog_str (app.py:577) using the real lane_to_str.
+    parts = ["t" + str(bpm)]
+    if bars:
+        parts.append("b" + str(bars))
+    if ramp:
+        parts.append("rmp%d/%d/%d" % (ramp.get("start", bpm), ramp["amt"], ramp["every"]))
+    if trainer:
+        parts.append("tr%d/%d" % (trainer["play"], trainer["mute"]))
+    for L in lanes:
+        parts.append(NS["lane_to_str"](L))
+    return ";".join(parts)
+
+
+def _gain_db(g):
+    if not g:
+        return 0
+    try:
+        return float(str(g).lstrip("@"))
+    except ValueError:
+        return 0
+
+
+def normalize(patch):
+    bpm, lanes, bars, ramp, trainer = NS["parse_program"](patch)
+    return {
+        "bpm": bpm,
+        "bars": bars,
+        "volume": None,   # device has no master-volume token
+        "countMs": 0,     # device has no count-in
+        "ramp": {"start": ramp.get("start", bpm), "amt": ramp["amt"], "every": ramp["every"]} if ramp else None,
+        "trainer": {"play": trainer["play"], "mute": trainer["mute"]} if trainer else None,
+        "rep": None,      # not parsed yet
+        "end": None,      # not parsed yet
+        "lanes": [
+            {
+                "sound": L["sound"],
+                "groups": list(L.get("groups", [4])),
+                "sub": L["sub"],
+                "swing": bool(L["swing"]),
+                "poly": bool(L["poly"]),
+                "mute": bool(L["mute"]),
+                "gainDb": _gain_db(L.get("gain", "")),
+                "levels": [int(v) for v in L["levels"]],
+            }
+            for L in lanes
+        ],
+    }
+
+
+def canonical(patch):
+    return _prog_str(*NS["parse_program"](patch))
+
+
+if __name__ == "__main__":
+    patch = sys.argv[1] if len(sys.argv) > 1 else ""
+    # Re-parse the canonical form too, so the runner can check idempotency in one call.
+    c1 = canonical(patch)
+    c2 = canonical(c1)
+    sys.stdout.write(json.dumps({"norm": normalize(patch), "canonical": c1, "idempotent": c1 == c2}))
diff --git a/tests/fixtures/track-format.json b/tests/fixtures/track-format.json
new file mode 100644
index 0000000..0ca56f1
--- /dev/null
+++ b/tests/fixtures/track-format.json
@@ -0,0 +1,183 @@
+{
+  "_comment": "Golden vectors for the PM track format. See docs/track-format.md. Each case: parse `in`, normalize, compare to `norm`. status: stable|divergence|new. expectFail lists impls (js/py) known to mismatch today; an UNlisted mismatch is a regression, and a listed PASS means the gap was fixed. NOTE: stable cases use EXPLICIT =patterns on every lane, because lanes with no =pattern currently produce DIFFERENT defaults on web vs device (see the 'default-pattern' case + docs §6).",
+  "cases": [
+    {
+      "id": "minimal",
+      "status": "stable",
+      "in": "t120;kick:4=Xxxx",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+    {
+      "id": "backbeat",
+      "status": "stable",
+      "in": "t120;kick:4=Xxxx;snare:4=.x.x;hatClosed:4/2=X.x.x.x.",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [
+          { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] },
+          { "sound": "snare", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [0,1,0,1] },
+          { "sound": "hatClosed", "groups": [4], "sub": 2, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,0,1,0,1,0,1,0] }
+        ] }
+    },
+    {
+      "id": "odd-meter-2+2+3",
+      "status": "stable",
+      "in": "t130;kick:2+2+3=x..x..x;hatClosed:2+2+3/2=X.x.X.x.X.x.x.",
+      "norm": { "bpm": 130, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [
+          { "sound": "kick", "groups": [2,2,3], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [1,0,0,1,0,0,1] },
+          { "sound": "hatClosed", "groups": [2,2,3], "sub": 2, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,0,1,0,2,0,1,0,2,0,1,0,1,0] }
+        ] }
+    },
+    {
+      "id": "swing",
+      "status": "stable",
+      "in": "t150;ride:4/2s=X.x.x.x.;kick:4=X..x",
+      "norm": { "bpm": 150, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [
+          { "sound": "ride", "groups": [4], "sub": 2, "swing": true, "poly": false, "mute": false, "gainDb": 0, "levels": [2,0,1,0,1,0,1,0] },
+          { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,0,0,1] }
+        ] }
+    },
+    {
+      "id": "ghost-notes",
+      "status": "stable",
+      "in": "t92;snare:4/3=..gg.gX.gg.g",
+      "norm": { "bpm": 92, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "snare", "groups": [4], "sub": 3, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [0,0,3,3,0,3,2,0,3,3,0,3] } ] }
+    },
+    {
+      "id": "polymeter",
+      "status": "stable",
+      "in": "t100;kick:4=Xxxx;claves:5=Xxxxx~",
+      "norm": { "bpm": 100, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [
+          { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] },
+          { "sound": "claves", "groups": [5], "sub": 1, "swing": false, "poly": true, "mute": false, "gainDb": 0, "levels": [2,1,1,1,1] }
+        ] }
+    },
+    {
+      "id": "disabled-lane",
+      "status": "stable",
+      "in": "t120;kick:4=Xxxx;hatClosed:4=Xxxx!",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [
+          { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] },
+          { "sound": "hatClosed", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": true, "gainDb": 0, "levels": [2,1,1,1] }
+        ] }
+    },
+    {
+      "id": "ramp",
+      "status": "stable",
+      "in": "t80;woodblock:4=Xxxx;rmp80/4/4",
+      "norm": { "bpm": 80, "bars": 0, "volume": null, "countMs": 0, "ramp": { "start": 80, "amt": 4, "every": 4 }, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "woodblock", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+    {
+      "id": "gap-trainer",
+      "status": "stable",
+      "in": "t100;kick:4=Xxxx;tr2/2",
+      "norm": { "bpm": 100, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": { "play": 2, "mute": 2 }, "rep": null, "end": null,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+    {
+      "id": "segment-bars",
+      "status": "stable",
+      "in": "t88;b8;kick:4=X.x.",
+      "norm": { "bpm": 88, "bars": 8, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,0,1,0] } ] }
+    },
+    {
+      "id": "gm-note-number-alias",
+      "status": "stable",
+      "in": "t120;36:4=Xxxx",
+      "note": "GM note 36 = kick. engine.js maps it; firmware keeps the numeric sound but plays it via SOUND_GM. Names should resolve the same.",
+      "expectFail": ["py"],
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+
+    {
+      "id": "default-pattern",
+      "status": "divergence",
+      "expectFail": ["py"],
+      "note": "DESIGN DECISION PENDING. A lane with no =pattern produces different defaults: web plays every subdivision (off-beats normal) and accents every beat -> [2,1,2,1...]; device rests off-beats -> [2,0,1,0...]. This affects shipped presets (e.g. 'Four-on-the-floor' hatClosed:4/2). Vector pins the web behavior as provisional reference; revise once the rule is chosen.",
+      "in": "t120;hatClosed:4/2",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "hatClosed", "groups": [4], "sub": 2, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,2,1,2,1,2,1] } ] }
+    },
+    {
+      "id": "euclid",
+      "status": "divergence",
+      "expectFail": ["py"],
+      "note": "Euclid (k,n) parsed by engine.js; firmware falls back to a plain bar. Spec = engine behavior.",
+      "in": "t120;kick:4(3,8)",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 2, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,0,0,1,0,0,1,0] } ] }
+    },
+    {
+      "id": "vol-and-countin",
+      "status": "divergence",
+      "expectFail": ["py"],
+      "note": "engine.js round-trips vol/cd; firmware drops them. Spec = preserve them.",
+      "in": "t120;vol80;cd2;kick:4=Xxxx",
+      "norm": { "bpm": 120, "bars": 0, "volume": 0.8, "countMs": 2000, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+    {
+      "id": "unknown-sound",
+      "status": "divergence",
+      "expectFail": ["py"],
+      "note": "engine.js maps unknown sounds to beep; firmware keeps the raw name. Spec = beep.",
+      "in": "t120;blorp:4=Xxxx",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "beep", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+    {
+      "id": "tempo-clamp-high",
+      "status": "divergence",
+      "expectFail": ["js"],
+      "note": "Firmware clamps t to [5,300]; engine.js does not. Spec = clamp everywhere.",
+      "in": "t999;kick:4=Xxxx",
+      "norm": { "bpm": 300, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+    {
+      "id": "empty-defaults-to-beep",
+      "status": "divergence",
+      "expectFail": ["js"],
+      "note": "Firmware injects a default beep:4; engine.js returns no lanes (editor guards emptiness). Spec = beep:4.",
+      "in": "",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": null, "end": null,
+        "lanes": [ { "sound": "beep", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+
+    {
+      "id": "end-stop",
+      "status": "new",
+      "expectFail": ["js", "py"],
+      "note": "Per-track playback flow — not yet implemented anywhere.",
+      "in": "t120;kick:4=Xxxx;end=stop",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": 1, "end": "stop",
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+    {
+      "id": "rep-then-next",
+      "status": "new",
+      "expectFail": ["js", "py"],
+      "note": "Play 4 cycles then auto-advance. next ⇒ +1.",
+      "in": "t120;kick:4=Xxxx;rep=4;end=next",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": 4, "end": 1,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    },
+    {
+      "id": "relative-goto-back-two",
+      "status": "new",
+      "expectFail": ["js", "py"],
+      "note": "Relative goto (D.S.): after 1 cycle, jump back two tracks.",
+      "in": "t120;kick:4=Xxxx;end=-2",
+      "norm": { "bpm": 120, "bars": 0, "volume": null, "countMs": 0, "ramp": null, "trainer": null, "rep": 1, "end": -2,
+        "lanes": [ { "sound": "kick", "groups": [4], "sub": 1, "swing": false, "poly": false, "mute": false, "gainDb": 0, "levels": [2,1,1,1] } ] }
+    }
+  ]
+}
diff --git a/tests/run.mjs b/tests/run.mjs
new file mode 100644
index 0000000..4062e7e
--- /dev/null
+++ b/tests/run.mjs
@@ -0,0 +1,97 @@
+#!/usr/bin/env node
+// Conformance runner for the PM track format.
+//   node tests/run.mjs            run all golden vectors against engine.js + app.py
+//   node tests/run.mjs -v         also print the expected/actual diff for every failure
+//
+// For each vector it parses `in` with both implementations, normalizes, and compares to
+// `norm`. A mismatch on an impl listed in the vector's `expectFail` is "known" (expected);
+// any other mismatch is a regression and fails the run. See docs/track-format.md.
+import { readFileSync } from "node:fs";
+import { fileURLToPath } from "node:url";
+import { dirname, join } from "node:path";
+import { execFileSync } from "node:child_process";
+import * as js from "./adapters/js_adapter.mjs";
+
+const here = dirname(fileURLToPath(import.meta.url));
+const verbose = process.argv.includes("-v");
+const fixtures = JSON.parse(readFileSync(join(here, "fixtures", "track-format.json"), "utf8"));
+const pyAdapter = join(here, "adapters", "py_adapter.py");
+
+// stable, key-sorted JSON so deep-equality is a string compare.
+const stable = (o) => JSON.stringify(o, (k, v) =>
+  v && typeof v === "object" && !Array.isArray(v)
+    ? Object.fromEntries(Object.keys(v).sort().map((kk) => [kk, v[kk]]))
+    : v);
+
+function runJs(patch) {
+  try {
+    return { norm: js.normalize(patch), canonical: js.canonical(patch), error: null };
+  } catch (e) {
+    return { norm: null, canonical: null, error: String(e.message || e) };
+  }
+}
+
+function runPy(patch) {
+  try {
+    const out = execFileSync("python3", [pyAdapter, patch], { encoding: "utf8" });
+    return { ...JSON.parse(out), error: null };
+  } catch (e) {
+    const msg = (e.stderr || "").toString().trim().split("\n").pop() || e.message;
+    return { norm: null, canonical: null, error: msg };
+  }
+}
+
+const want = stable; // alias
+let regressions = 0, fixedNowCount = 0, nonIdempotent = 0;
+const rows = [];
+
+function jsIdempotent(patch) {
+  try { const c1 = js.canonical(patch); return c1 === js.canonical(c1); } catch { return false; }
+}
+
+for (const c of fixtures.cases) {
+  const expected = want(c.norm);
+  const expectFail = new Set(c.expectFail || []);
+  const r = { id: c.id, status: c.status };
+
+  for (const [impl, res] of [["js", runJs(c.in)], ["py", runPy(c.in)]]) {
+    // serialize(parse(x)) must be stable under re-parsing (no silent drift on round-trip).
+    const idem = impl === "js" ? jsIdempotent(c.in) : res.idempotent !== false;
+    if (res.error == null && !idem) { nonIdempotent++; console.log(`  ! non-idempotent serialize: ${c.id} [${impl}]`); }
+    const ok = res.error == null && want(res.norm) === expected;
+    const known = expectFail.has(impl);
+    let mark;
+    if (ok && !known) mark = "PASS";
+    else if (ok && known) { mark = "FIXED"; fixedNowCount++; }       // listed as failing but now passes
+    else if (!ok && known) mark = "known";                          // expected divergence/not-built
+    else { mark = "FAIL"; regressions++; }                          // unexpected → regression
+    r[impl] = mark;
+    r[impl + "_res"] = res;
+    if (mark === "FAIL" && verbose) {
+      console.log(`\n--- ${c.id} [${impl}] expected vs actual ---`);
+      console.log("expected:", expected);
+      console.log("actual:  ", res.error ? "ERROR " + res.error : want(res.norm));
+    }
+  }
+  rows.push(r);
+}
+
+// ---- report ----
+const pad = (s, n) => String(s).padEnd(n);
+console.log("\n  PM track-format conformance\n");
+console.log("  " + pad("case", 26) + pad("status", 13) + pad("engine.js", 11) + "app.py");
+console.log("  " + "-".repeat(58));
+const glyph = { PASS: "✓ pass", known: "· known", FAIL: "✗ FAIL", FIXED: "★ fixed" };
+for (const r of rows) {
+  console.log("  " + pad(r.id, 26) + pad(r.status, 13) + pad(glyph[r.js], 11) + glyph[r.py]);
+}
+
+const counts = rows.reduce((a, r) => { a[r.js] = (a[r.js] || 0) + 1; a[r.py] = (a[r.py] || 0) + 1; return a; }, {});
+console.log("\n  " + Object.entries(counts).map(([k, v]) => `${glyph[k] || k}: ${v}`).join("   "));
+if (fixedNowCount) console.log(`\n  ${fixedNowCount} case(s) marked expectFail now PASS — update the fixture (remove them from expectFail).`);
+if (nonIdempotent) console.log(`  ${nonIdempotent} non-idempotent serialize(s) above.`);
+if (regressions || nonIdempotent) {
+  console.log(`\n  ✗ ${regressions} unexpected failure(s), ${nonIdempotent} round-trip issue(s). Run with -v for diffs.\n`);
+  process.exit(1);
+}
+console.log("\n  ✓ no unexpected failures; serialize round-trips are stable.\n");