tf-mediapipe getModelTask has no single-flight: two concurrent callers both miss the cache and both build a GPU task

[sroussey]

Defect

providers/tf-mediapipe/src/ai/common/TFMP_Runtime.ts caches built MediaPipe task instances by model_path + matched options, but the cache check and the insertion are separated by an await, with nothing recording the in-flight build.

TFMP_Runtime.ts:238-246:

const cachedTasks = modelTaskCache.get(model_path);
if (cachedTasks) {
  const matchedTask = cachedTasks.find((cached) => optionsMatch(cached.options, lookupOptions));
  if (matchedTask) {
    return matchedTask.task;
  }
}

const wasmFileset = await getWasmTask(model, emit, signal);   // ← yields

…and the result is only published at :271:

modelTaskCache.get(model_path)!.push(cachedTask);

Two concurrent callers with the same model_path and matching options both reach :238 before either reaches :271. Both miss, both download/instantiate, and one of the two built tasks is orphaned in the cache list — holding a WebGPU/WASM handle nothing will ever dispose.

Evidence

• providers/tf-mediapipe/src/ai/common/TFMP_Runtime.ts:238-246 — check
• providers/tf-mediapipe/src/ai/common/TFMP_Runtime.ts:271 — publish
• No in-flight map: grep -n "InFlight\|inFlight\|pending" providers/tf-mediapipe/src/ai/common/TFMP_Runtime.ts finds nothing on this path.

The package already contains the correct pattern for its sibling runtime — TFMP_GenaiRuntime.ts uses withGenaiLock / isGenaiBusy — and the cactus provider has the shape too (cactusEngineLoadsInFlight, referenced in providers/cactus/src/ai.ts's export comment). This is a copy of an in-tree pattern, not a new design.

Proposed fix

Add a Map<string, Promise<TaskInstance>> keyed the same way the cache is, populated before the first await and cleared in finally; return the shared promise on a hit. ~15 LOC.

Why it matters

Every vision run-fn in the package goes through getModelTask, and a fanned-out graph (MapTask over N images) hits it N times concurrently by construction. On a WebGPU delegate the duplicate build is a second GPU context and a second copy of the model weights, with the loser leaked for the process lifetime.

Carried as a review finding for five cycles (2026-08-03 onward) without being tracked. The 2026-08-31 window touched nine other files in this package (388a2b8, 9ada0ea) and not this one. Filing it. Snapshot: workglow-dev/prd → analysis/grades/2026-08-31/libs-providers.md.