@openparachute/vault 0.6.5-rc.9 → 0.7.0-rc.1

package/src/transcription/build.test.ts

@@ -6,6 +6,9 @@ import {
   detectCompiler,
   compilerInstallHint,
   cliSourceUrl,
+  patchMainCppBackendInit,
+  BACKEND_INIT_PATCH_ANCHOR,
+  BACKEND_INIT_PATCH_MARKER,
   CLI_SOURCE_FILES,
   CXX_CANDIDATES,
   TRANSCRIBE_CPP_SOURCE_REF,
@@ -17,10 +20,26 @@ import {
  * transcribe-cli build tests (scribe-fold Phase 2b). Pure recipe shape +
  * fully-mocked orchestration — NO network, NO real compiler (the real compile
  * is exercised only by the install-time live verify). Covers the toolchain-
- * absent, fetch-failure, compile-failure, and success branches, and the
- * "never leave a half-built binary" invariant.
+ * absent, fetch-failure, patch-failure, compile-failure, and success branches,
+ * and the "never leave a half-built binary" invariant.
  */
 
+/** A minimal stand-in for the pinned main.cpp: log-sink install + the batch-
+ *  mode comment the vault#534 patch anchors on. */
+const MAIN_CPP_FIXTURE = [
+  "int main(int argc, char ** argv) {",
+  "    if (!args.quiet) {",
+  "        transcribe_log_set(log_cb, nullptr);",
+  "    }",
+  "",
+  BACKEND_INIT_PATCH_ANCHOR,
+  "    // the model ONCE and reuses the context across all files.",
+  "    if (!args.batch_file.empty()) {",
+  "    }",
+  "}",
+  "",
+].join("\n");
+
 describe("detectCompiler", () => {
   test("returns the first candidate found (c++ preferred)", () => {
     const which = (c: string) => (c === "c++" ? "/usr/bin/c++" : null);
@@ -106,15 +125,31 @@ describe("buildCompileCommand — the proven recipe shape", () => {
 
 /** Build a deps object with call-recording mocks; overrides win. */
 function mkDeps(over: Partial<BuildCliDeps> = {}): BuildCliDeps & {
-  calls: { fetch: number; compile: number; removed: string[]; renamed: Array<[string, string]> };
+  calls: {
+    fetch: number;
+    compile: number;
+    removed: string[];
+    renamed: Array<[string, string]>;
+    written: Array<[string, string]>;
+  };
 } {
-  const calls = { fetch: 0, compile: 0, removed: [] as string[], renamed: [] as Array<[string, string]> };
+  const calls = {
+    fetch: 0,
+    compile: 0,
+    removed: [] as string[],
+    renamed: [] as Array<[string, string]>,
+    written: [] as Array<[string, string]>,
+  };
   const deps: BuildCliDeps = {
     platform: "darwin",
     which: (c) => (c === "c++" ? "/usr/bin/c++" : null),
     fetchSource: async () => {
       calls.fetch++;
     },
+    readFile: () => MAIN_CPP_FIXTURE,
+    writeFile: (p, content) => {
+      calls.written.push([p, content]);
+    },
     compile: async (): Promise<CompileResult> => {
       calls.compile++;
       return { exitCode: 0, stderr: "" };
@@ -134,8 +169,36 @@ function mkDeps(over: Partial<BuildCliDeps> = {}): BuildCliDeps & {
 const INPUT = { srcDir: "/t/.src", libsDir: "/t/libs", binPath: "/t/bin/transcribe-cli" };
 const TMP = `${INPUT.binPath}.building`; // the temp output the build compiles to
 
+describe("patchMainCppBackendInit — the vault#534 backend-init source patch", () => {
+  test("injects transcribe_init_backends_default() before the batch-mode anchor", () => {
+    const patched = patchMainCppBackendInit(MAIN_CPP_FIXTURE);
+    expect(patched).toContain("transcribe_init_backends_default();");
+    expect(patched).toContain(BACKEND_INIT_PATCH_MARKER);
+    // The init call lands BEFORE the inference paths (the anchor) and AFTER
+    // the log-sink install, so backend plugin-load logs route through the sink.
+    const initAt = patched.indexOf("transcribe_init_backends_default();");
+    expect(initAt).toBeGreaterThan(patched.indexOf("transcribe_log_set(log_cb, nullptr);"));
+    expect(initAt).toBeLessThan(patched.indexOf(BACKEND_INIT_PATCH_ANCHOR));
+    // Everything else is preserved verbatim: removing the injected block
+    // restores the original source exactly.
+    const [before, after] = patched.split(/    \/\/ \[parachute-vault[\s\S]*?transcribe_init_backends_default\(\);\n\n/);
+    expect(`${before}${after}`).toBe(MAIN_CPP_FIXTURE);
+  });
+
+  test("idempotent: an already-patched source is returned unchanged", () => {
+    const once = patchMainCppBackendInit(MAIN_CPP_FIXTURE);
+    expect(patchMainCppBackendInit(once)).toBe(once);
+  });
+
+  test("anchor missing ⇒ throws a loud, actionable error naming the pin + vault#534", () => {
+    expect(() => patchMainCppBackendInit("int main() { return 0; }\n")).toThrow(
+      /anchor not found[\s\S]*vault#534/,
+    );
+  });
+});
+
 describe("buildTranscribeCli — orchestration branches", () => {
-  test("SUCCESS: fetch → compile ok → temp exists ⇒ promote (rename) + { ok:true }", async () => {
+  test("SUCCESS: fetch → patch → compile ok → temp exists ⇒ promote (rename) + { ok:true }", async () => {
     const deps = mkDeps();
     const r = await buildTranscribeCli(INPUT, deps);
     expect(r.ok).toBe(true);
@@ -149,11 +212,29 @@ describe("buildTranscribeCli — orchestration branches", () => {
     }
     expect(deps.calls.fetch).toBe(1);
     expect(deps.calls.compile).toBe(1);
+    // the vault#534 backend-init patch was written into the fetched main.cpp
+    expect(deps.calls.written).toHaveLength(1);
+    const [patchedPath, patchedContent] = deps.calls.written[0]!;
+    expect(patchedPath).toBe(join(INPUT.srcDir, "examples", "cli", "main.cpp"));
+    expect(patchedContent).toContain("transcribe_init_backends_default();");
     // only the TEMP is cleared pre-compile; the fresh temp is renamed into place
     expect(deps.calls.removed).toEqual([TMP]);
     expect(deps.calls.renamed).toEqual([[TMP, INPUT.binPath]]);
   });
 
+  test("PATCH FAILURE: anchor missing in fetched source ⇒ { ok:false, stage:'patch' }, no write, no compile", async () => {
+    const deps = mkDeps({ readFile: () => "int main() { return 0; }\n" });
+    const r = await buildTranscribeCli(INPUT, deps);
+    expect(r.ok).toBe(false);
+    if (!r.ok) {
+      expect(r.stage).toBe("patch");
+      expect(r.message).toContain("anchor not found");
+      expect(r.message).toContain("vault#534");
+    }
+    expect(deps.calls.written).toEqual([]);
+    expect(deps.calls.compile).toBe(0);
+  });
+
   test("TOOLCHAIN ABSENT: no compiler ⇒ { ok:false, stage:'toolchain' }, no fetch/compile", async () => {
     const deps = mkDeps({ which: () => null });
     const r = await buildTranscribeCli(INPUT, deps);

package/src/transcription/build.ts

@@ -6,8 +6,10 @@
  * `install.ts` + the provider-seam design doc. This module closes that gap: it
  * fetches the CLI's source (`examples/cli/main.cpp` + `examples/common/wav.cpp`
  * + the public headers) from the transcribe.cpp repo pinned at the v0.1.1 tag
- * commit, then compiles it with the host C++ compiler against the extracted
- * prebuilt dylibs. The result is a ~127KB driver that links the prebuilt
+ * commit, applies the **vault#534 backend-init patch** (see
+ * `patchMainCppBackendInit` below — temporary until upstream ships the fix),
+ * then compiles it with the host C++ compiler against the extracted prebuilt
+ * dylibs. The result is a ~127KB driver that links the prebuilt
  * `libtranscribe` at runtime — **no ggml / Metal rebuild**.
  *
  * ## The proven recipe (live-verified 2026-07-03, macOS arm64 M4)
@@ -34,10 +36,11 @@
  * ## Testability
  *
  * `buildTranscribeCli` is a pure orchestrator over injected side effects
- * (`which` / `fetchSource` / `compile` / `exists` / `removeBin`), mirroring the
- * provider's `SpawnRunner` seam. Tests exercise every branch — toolchain
- * absent, fetch failure, compile failure, success — with no network and no real
- * compiler. The real compile is exercised only by the install-time live verify.
+ * (`which` / `fetchSource` / `readFile` / `writeFile` / `compile` / `exists` /
+ * `removeBin`), mirroring the provider's `SpawnRunner` seam. Tests exercise
+ * every branch — toolchain absent, fetch failure, patch failure, compile
+ * failure, success — with no network and no real compiler. The real compile is
+ * exercised only by the install-time live verify.
  */
 
 import { join } from "path";
@@ -78,6 +81,63 @@ export function cliSourceUrl(repoPath: string): string {
   return `${RAW_BASE}/${repoPath}`;
 }
 
+// ---------------------------------------------------------------------------
+// Build-time source patch: unconditional backend init (vault#534 blocker 2)
+// ---------------------------------------------------------------------------
+
+/** Marker comment baked into the injected code — makes the patch idempotent
+ *  and greppable in a fetched source tree. */
+export const BACKEND_INIT_PATCH_MARKER = "[parachute-vault vault#534 backend-init patch]";
+
+/**
+ * The anchor line in the pinned `examples/cli/main.cpp` the patch injects
+ * before: the batch-mode comment that opens the inference half of `main()`,
+ * directly AFTER the log-sink install (so backend plugin-load logs route
+ * through the CLI's log callback) and after the `--list-devices` early return.
+ * Unique at the pinned ref (verified); if the pin ever advances and this
+ * anchor drifts, the build fails LOUDLY (stage "patch") so the patch gets
+ * re-evaluated against the new source instead of silently shipping a CLI that
+ * can't transcribe on Linux.
+ */
+export const BACKEND_INIT_PATCH_ANCHOR =
+  "    // Batch mode: --batch reads a file list, one wav path per line. Loads";
+
+const BACKEND_INIT_PATCH_BLOCK = `    // ${BACKEND_INIT_PATCH_MARKER}
+    // Upstream (at the pinned ${TRANSCRIBE_CPP_SOURCE_REF.slice(0, 12)}) only calls
+    // transcribe_init_backends_default() in the --list-devices path. On Linux
+    // the CPU backends are dlopen plugins (libggml-cpu-*.so), so with no init
+    // the ggml device registry is empty and EVERY transcription fails
+    // ("whisper: failed to initialize CPU backend", exit 1). macOS is masked:
+    // its libtranscribe.dylib direct-links libggml-cpu, which self-registers
+    // at library load. Init the default backends unconditionally before any
+    // inference (batch or single-file). Verified fix on real Linux containers,
+    // both arches — see vault#534. DROP THIS PATCH when upstream ships the fix
+    // and the source pin advances past it.
+    transcribe_init_backends_default();
+
+`;
+
+/**
+ * Apply the vault#534 backend-init patch to the fetched `main.cpp` source.
+ * Pure (string → string) so tests pin the behavior without a filesystem.
+ * Idempotent: an already-patched source is returned unchanged. Throws with a
+ * clear operator-facing message when the anchor is missing — the caller
+ * surfaces it as a `stage: "patch"` build failure (fail loudly; never compile
+ * an unpatched CLI on the assumption it'll work).
+ */
+export function patchMainCppBackendInit(source: string): string {
+  if (source.includes(BACKEND_INIT_PATCH_MARKER)) return source;
+  const idx = source.indexOf(BACKEND_INIT_PATCH_ANCHOR);
+  if (idx === -1) {
+    throw new Error(
+      `backend-init patch anchor not found in examples/cli/main.cpp at ${TRANSCRIBE_CPP_SOURCE_REF.slice(0, 12)} — ` +
+        `the pinned source no longer matches the vault#534 patch. If the pin advanced, check whether upstream ` +
+        `now calls transcribe_init_backends_default() unconditionally (then drop the patch); otherwise re-anchor it.`,
+    );
+  }
+  return source.slice(0, idx) + BACKEND_INIT_PATCH_BLOCK + source.slice(idx);
+}
+
 /** C++ compiler binaries we probe for, in preference order. */
 export const CXX_CANDIDATES = ["c++", "clang++", "g++"] as const;
 
@@ -158,6 +218,10 @@ export interface BuildCliDeps {
   which: (cmd: string) => string | null | undefined;
   /** Fetch `files` (repo-relative) into `srcDir`. Throws on network failure. */
   fetchSource: (files: readonly string[], srcDir: string) => Promise<void>;
+  /** Read a fetched source file as UTF-8 (production: `fs.readFileSync`). */
+  readFile: (p: string) => string;
+  /** Write a patched source file (production: `fs.writeFileSync`). */
+  writeFile: (p: string, content: string) => void;
   /** Run the compiler argv. */
   compile: (cmd: string[]) => CompileResult | Promise<CompileResult>;
   /** Existence probe (production: `fs.existsSync`). */
@@ -185,7 +249,7 @@ export type CliBuildResult =
   | {
       ok: false;
       /** Which stage failed — drives the operator guidance. */
-      stage: "toolchain" | "fetch" | "compile";
+      stage: "toolchain" | "fetch" | "patch" | "compile";
       message: string;
       compiler?: string;
       /** The compile argv (as a string) when we got far enough to build one. */
@@ -223,6 +287,22 @@ export async function buildTranscribeCli(
     };
   }
 
+  // Apply the vault#534 backend-init source patch BEFORE compiling — without
+  // it the built CLI exits 1 on every transcription on Linux (dlopen'd CPU
+  // backends are never registered). A missing anchor is a loud, typed failure,
+  // never a silent skip.
+  const mainCppPath = join(input.srcDir, "examples", "cli", "main.cpp");
+  try {
+    deps.writeFile(mainCppPath, patchMainCppBackendInit(deps.readFile(mainCppPath)));
+  } catch (err) {
+    return {
+      ok: false,
+      stage: "patch",
+      compiler,
+      message: err instanceof Error ? err.message : String(err),
+    };
+  }
+
   const common = { platform: deps.platform, compiler, srcDir: input.srcDir, libsDir: input.libsDir };
   // The reported command targets the FINAL binPath (what an operator would run
   // by hand to reproduce). The actual compile writes to a temp sibling and is

package/src/transcription/capability.test.ts

@@ -90,4 +90,29 @@ describe("defaultTranscriptionProvider — provider selection", () => {
     process.env.TRANSCRIPTION_PROVIDER = "transcribe-cpp";
     expect(defaultTranscriptionProvider().name).toBe("transcribe-cpp");
   });
+
+  test("TRANSCRIPTION_PROVIDER=parakeet-mlx → the parakeet-mlx provider (2b)", () => {
+    process.env.TRANSCRIPTION_PROVIDER = "parakeet-mlx";
+    expect(defaultTranscriptionProvider().name).toBe("parakeet-mlx");
+  });
+
+  test("TRANSCRIPTION_PROVIDER=onnx-asr → the onnx-asr provider (2b)", () => {
+    process.env.TRANSCRIPTION_PROVIDER = "onnx-asr";
+    expect(defaultTranscriptionProvider().name).toBe("onnx-asr");
+  });
+
+  test("an unconfigured python provider resolves to disabled (no throw)", async () => {
+    // No venv/PATH binary in a fresh env → available() is ok:false and the
+    // landing capability is {enabled:false} rather than a crash.
+    process.env.TRANSCRIPTION_PROVIDER = "parakeet-mlx";
+    const prevBin = process.env.PARAKEET_MLX_BIN;
+    process.env.PARAKEET_MLX_BIN = "/definitely/not/a/real/parakeet-mlx";
+    try {
+      const cap = await resolveTranscriptionCapability(defaultTranscriptionProvider());
+      expect(cap.enabled).toBe(false);
+    } finally {
+      if (prevBin === undefined) delete process.env.PARAKEET_MLX_BIN;
+      else process.env.PARAKEET_MLX_BIN = prevBin;
+    }
+  });
 });

package/src/transcription/capability.ts

@@ -17,9 +17,18 @@
 import type { TranscriptionProvider } from "../../core/src/transcription/provider.ts";
 import { ScribeHttpProvider } from "./providers/scribe-http.ts";
 import { TranscribeCppProvider } from "./providers/transcribe-cpp.ts";
+import { ParakeetMlxProvider } from "./providers/parakeet-mlx.ts";
+import { OnnxAsrProvider } from "./providers/onnx-asr.ts";
 import { getCachedScribeUrl } from "../scribe-discovery.ts";
 import { resolveScribeAuthToken } from "../scribe-env.ts";
-import { resolveTranscriptionProviderName, resolveTranscribeCppPaths } from "./select.ts";
+import {
+  resolveTranscriptionProviderName,
+  resolveTranscribeCppPaths,
+  resolveParakeetMlxBin,
+  resolveParakeetMlxModel,
+  resolveOnnxAsrBin,
+  resolveOnnxAsrModel,
+} from "./select.ts";
 
 export interface TranscriptionCapability {
   /** True when a provider is configured AND available. Notes gates the mic on this. */
@@ -36,16 +45,32 @@ export interface TranscriptionCapability {
  *   - `transcribe-cpp` → the local provider, resolving the installed binary +
  *     GGUF model paths; `available()` is `false` until `transcription install`
  *     has run.
+ *   - `parakeet-mlx` / `onnx-asr` (scribe-fold Phase 2b) → the Python-based
+ *     local providers, resolving the binary via env override → managed venv →
+ *     PATH; `available()` is `false` until a runnable binary exists.
  *   - `scribe-http` (default) → the remote provider (URL from
  *     `SCRIBE_URL`/services.json, bearer from `SCRIBE_AUTH_TOKEN`). When scribe
  *     isn't discoverable the URL is undefined and it reports itself unavailable
  *     rather than throwing.
  */
 export function defaultTranscriptionProvider(): TranscriptionProvider {
-  if (resolveTranscriptionProviderName() === "transcribe-cpp") {
+  const name = resolveTranscriptionProviderName();
+  if (name === "transcribe-cpp") {
     const paths = resolveTranscribeCppPaths();
     return new TranscribeCppProvider({ binPath: paths.binPath, modelPath: paths.modelPath });
   }
+  if (name === "parakeet-mlx") {
+    return new ParakeetMlxProvider({
+      binPath: resolveParakeetMlxBin(),
+      model: resolveParakeetMlxModel(),
+    });
+  }
+  if (name === "onnx-asr") {
+    return new OnnxAsrProvider({
+      binPath: resolveOnnxAsrBin(),
+      model: resolveOnnxAsrModel(),
+    });
+  }
   return new ScribeHttpProvider({
     url: getCachedScribeUrl(),
     token: resolveScribeAuthToken(),

package/src/transcription/download.test.ts

@@ -0,0 +1,101 @@
+import { describe, test, expect, afterAll } from "bun:test";
+import { existsSync, mkdtempSync, readFileSync, rmSync } from "fs";
+import { tmpdir } from "os";
+import { join } from "path";
+import { downloadTo } from "./download.ts";
+
+/**
+ * `downloadTo` tests (vault#534 blocker 1). The manual body→FileSink pump
+ * replaced `Bun.write(dest, resp)`, which hangs forever on Linux for large
+ * responses. These run the pump against an in-process `Bun.serve` (no
+ * subprocess involved — see CLAUDE.md's spawnSync note; plain in-process
+ * fetches are fine) and pin: bytes land intact, HTTP errors throw, and a
+ * mid-stream failure never leaves a partial file behind.
+ */
+
+const dir = mkdtempSync(join(tmpdir(), "vault-download-test-"));
+afterAll(() => rmSync(dir, { recursive: true, force: true }));
+
+// 4MB of deterministic non-trivial bytes — large enough to stream in many
+// chunks (the hang was specific to the large-response path).
+const PAYLOAD = new Uint8Array(4 * 1024 * 1024);
+for (let i = 0; i < PAYLOAD.length; i++) PAYLOAD[i] = (i * 31 + (i >> 8)) & 0xff;
+
+function serve(handler: (req: Request) => Response | Promise<Response>) {
+  const server = Bun.serve({ port: 0, fetch: handler });
+  return { server, url: `http://127.0.0.1:${server.port}` };
+}
+
+describe("downloadTo", () => {
+  test("streams a multi-MB body to disk byte-for-byte", async () => {
+    const { server, url } = serve(() => new Response(PAYLOAD));
+    const dest = join(dir, "ok.bin");
+    try {
+      await downloadTo(`${url}/file`, dest);
+      const got = readFileSync(dest);
+      expect(got.byteLength).toBe(PAYLOAD.byteLength);
+      expect(Buffer.from(got).equals(Buffer.from(PAYLOAD))).toBe(true);
+    } finally {
+      server.stop(true);
+    }
+  });
+
+  test("non-OK status throws with the status + URL, writes nothing", async () => {
+    const { server, url } = serve(() => new Response("nope", { status: 404 }));
+    const dest = join(dir, "missing.bin");
+    try {
+      expect(downloadTo(`${url}/gone`, dest)).rejects.toThrow(/download failed \(404\)/);
+      expect(existsSync(dest)).toBe(false);
+    } finally {
+      server.stop(true);
+    }
+  });
+
+  test("truncated download throws AND removes the partial file", async () => {
+    // Raw TCP fixture: a real content-length response (how GitHub releases +
+    // HuggingFace serve these binaries) whose connection dies mid-body.
+    // Bun.serve can't fixture this — it re-frames streamed responses as
+    // chunked (dropping content-length), and Bun's fetch ends a truncated
+    // chunked stream CLEANLY, so truncation is only detectable on the
+    // content-length path `downloadTo` verifies.
+    const partial = PAYLOAD.slice(0, 64 * 1024);
+    const listener = Bun.listen({
+      hostname: "127.0.0.1",
+      port: 0,
+      socket: {
+        data(socket) {
+          socket.write(
+            `HTTP/1.1 200 OK\r\ncontent-type: application/octet-stream\r\ncontent-length: ${PAYLOAD.byteLength}\r\n\r\n`,
+          );
+          socket.write(partial);
+          socket.end(); // die after 64KB of a declared 4MB
+        },
+        open() {},
+        close() {},
+        error() {},
+      },
+    });
+    const dest = join(dir, "partial.bin");
+    try {
+      let threw = false;
+      try {
+        await downloadTo(`http://127.0.0.1:${listener.port}/flaky`, dest);
+      } catch (err) {
+        threw = true;
+        // The truncation surfaces differently per platform: on macOS the body
+        // stream errors mid-pump (wrapped "failed after N bytes") or the
+        // short read trips the content-length check ("truncated"); on Linux
+        // Bun's fetch() itself rejects on the early socket close. All are
+        // loud failures — the invariants are THROWS + NO PARTIAL FILE.
+        expect(String(err)).toMatch(
+          /failed after \d+ bytes|truncated|socket connection was closed|connection closed/i,
+        );
+      }
+      expect(threw).toBe(true);
+      // The half-written file must not survive to be mistaken for a good artifact.
+      expect(existsSync(dest)).toBe(false);
+    } finally {
+      listener.stop(true);
+    }
+  });
+});

package/src/transcription/download.ts

@@ -0,0 +1,71 @@
+/**
+ * Streaming file download for `transcription install` (tarball, GGUF models,
+ * CLI sources).
+ *
+ * ## Why a manual pump instead of `Bun.write(dest, resp)` (vault#534 blocker 1)
+ *
+ * `Bun.write(dest, response)` HANGS FOREVER on Linux when streaming a large
+ * Response body to disk — zero bytes land, no error, no timeout. Reproduced
+ * deterministically in real Linux containers on BOTH arches (aarch64 native +
+ * x86_64 emulated) across bun 1.2.23 / 1.3.13 / 1.3.14, while the identical
+ * code works on macOS. In the same environment `curl` pulls the same URL at
+ * full speed and `await resp.arrayBuffer()` gets all 26MB in ~2s — the bug is
+ * specifically Bun's Response→file streaming fast path. Manually pumping
+ * `resp.body` chunks into a `Bun.file(dest).writer()` moves the same 26MB in
+ * <1s and works on both platforms, so that's what we do everywhere (no
+ * platform gate). See vault#534 for the full container verification.
+ */
+
+import { rmSync } from "fs";
+
+/**
+ * Download `url` to `dest` (follows redirects). Streams chunk-by-chunk — never
+ * buffers the whole body (models run up to ~660MB). On any mid-stream failure
+ * the partial file is removed so a retry never trusts a truncated artifact.
+ * When the server sent an honest `content-length` (no content-encoding
+ * transform), a byte-count mismatch is treated as a failed download too.
+ */
+export async function downloadTo(url: string, dest: string): Promise<void> {
+  const resp = await fetch(url, { redirect: "follow" });
+  if (!resp.ok) throw new Error(`download failed (${resp.status}) for ${url}`);
+  if (!resp.body) throw new Error(`download failed (empty response body) for ${url}`);
+
+  // content-length is only the on-the-wire byte count when no transfer
+  // decompression happened; fetch auto-decodes content-encoding'd bodies, so
+  // only enforce the size check for identity responses (the normal case for
+  // release tarballs + GGUFs — both already-compressed binary).
+  const encoding = resp.headers.get("content-encoding");
+  const lengthHeader = resp.headers.get("content-length");
+  const expectedBytes =
+    (!encoding || encoding === "identity") && lengthHeader && /^\d+$/.test(lengthHeader)
+      ? Number(lengthHeader)
+      : null;
+
+  const sink = Bun.file(dest).writer();
+  let written = 0;
+  try {
+    for await (const chunk of resp.body) {
+      sink.write(chunk);
+      written += chunk.byteLength;
+    }
+    await sink.end();
+  } catch (err) {
+    // Flush/close what we can, then remove the partial file — a half-written
+    // tarball/model must never survive to be mistaken for a good artifact.
+    try {
+      await sink.end();
+    } catch {
+      // best-effort close; the rm below is what matters
+    }
+    rmSync(dest, { force: true });
+    const msg = err instanceof Error ? err.message : String(err);
+    throw new Error(`download of ${url} failed after ${written} bytes: ${msg}`);
+  }
+
+  if (expectedBytes !== null && written !== expectedBytes) {
+    rmSync(dest, { force: true });
+    throw new Error(
+      `download of ${url} was truncated: got ${written} bytes, expected ${expectedBytes} (content-length)`,
+    );
+  }
+}