From 856999559179001cbff4b715999ceb2dec329203 Mon Sep 17 00:00:00 2001
From: tzhouam <tzhouam@connect.ust.hk>
Date: Wed, 23 Sep 2026 20:52:43 +0800
Subject: [PATCH] Add bounded Omni Qwen3-TTS CPU Radeon whole-session stage

---
 .../experiments/probe_omni_crisp_tts.py       | 180 +++++++
 .../probe_omni_crisp_tts_entrypoint.py        | 110 ++++
 .../results/e2e_expansion_20260923/README.md  |   3 +-
 .../evidence/qwen_tts_omni_hybrid/README.md   |  28 +
 ...p_tts_pinned_profile20_20260923.driver.log | 495 ++++++++++++++++++
 ...i_crisp_tts_pinned_profile20_20260923.json | 491 +++++++++++++++++
 ...ni_crisp_tts_pinned_profile20_20260923.log | 167 ++++++
 ...isp_tts_pinned_profile20_asr_20260923.json |  16 +
 ...risp_tts_pinned_profile20_asr_20260923.log |  17 +
 ...risp_tts_pinned_public_20260923.driver.log |  52 ++
 ...omni_crisp_tts_pinned_public_20260923.json |  26 +
 .../omni_crisp_tts_pinned_public_20260923.log |  79 +++
 ...s_pinned_reject_device_20260923.driver.log | 100 ++++
 ...isp_tts_pinned_reject_device_20260923.json |  26 +
 ...s_pinned_reject_memory_20260923.driver.log | 100 ++++
 ...isp_tts_pinned_reject_memory_20260923.json |  26 +
 .../qwen_tts_radeon890m_joint/README.md       |   2 +-
 .../evidence/summary/README.md                |   6 +-
 docs/design/omni_edge_backends.md             |   9 +-
 docs/design/omni_edge_e2e_roadmap.md          |   2 +
 vllm_omni/config/pipeline_registry.py         |   7 +-
 vllm_omni/engine/backends/__init__.py         |   4 +
 vllm_omni/engine/backends/crisp_tts.py        | 352 +++++++++++++
 .../models/qwen3_tts/pipeline.py              |  17 +
 24 files changed, 2308 insertions(+), 7 deletions(-)
 create mode 100644 benchmarks/edge_harness/experiments/probe_omni_crisp_tts.py
 create mode 100644 benchmarks/edge_harness/experiments/probe_omni_crisp_tts_entrypoint.py
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.driver.log
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.json
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.log
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_asr_20260923.json
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_asr_20260923.log
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.driver.log
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.json
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.log
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_device_20260923.driver.log
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_device_20260923.json
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_memory_20260923.driver.log
 create mode 100644 benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_memory_20260923.json
 create mode 100644 vllm_omni/engine/backends/crisp_tts.py

diff --git a/benchmarks/edge_harness/experiments/probe_omni_crisp_tts.py b/benchmarks/edge_harness/experiments/probe_omni_crisp_tts.py
new file mode 100644
index 000000000..ee55089f5
--- /dev/null
+++ b/benchmarks/edge_harness/experiments/probe_omni_crisp_tts.py
@@ -0,0 +1,180 @@
+#!/usr/bin/env python3
+"""Profile pinned Qwen3-TTS CustomVoice through the Omni CrispASR stage."""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import hashlib
+import json
+import math
+import time
+import wave
+from pathlib import Path
+from types import SimpleNamespace
+
+import numpy as np
+
+
+def _rank(values: list[float], fraction: float) -> float:
+    return sorted(values)[math.ceil(len(values) * fraction) - 1]
+
+
+def _save_wav(path: Path, audio) -> None:
+    pcm = np.rint(audio.numpy() * 32768).astype("<i2").tobytes()
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with wave.open(str(path), "wb") as wav:
+        wav.setnchannels(1)
+        wav.setsampwidth(2)
+        wav.setframerate(24000)
+        wav.writeframes(pcm)
+
+
+async def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    for name in (
+        "server_bin", "server_sha256", "talker_file", "talker_sha256", "codec_file",
+        "codec_sha256", "server_log", "output_report", "output_wav",
+        "punc_file", "punc_sha256",
+    ):
+        parser.add_argument("--" + name.replace("_", "-"), required=True)
+    parser.add_argument("--expected-gpu-name", default="AMD Radeon(TM) 890M Graphics")
+    parser.add_argument("--warmups", type=int, default=1)
+    parser.add_argument("--repeats", type=int, default=20)
+    parser.add_argument("--abort-check", action="store_true")
+    parser.add_argument("--capacity-gib", type=int, default=16)
+    parser.add_argument("--reserve-gib", type=int, default=8)
+    args = parser.parse_args()
+    if args.warmups < 0 or args.repeats <= 0 or args.reserve_gib <= 0 or args.capacity_gib < args.reserve_gib:
+        parser.error("invalid warmup/repeat count or memory budget")
+
+    from vllm_omni.config.stage_config import DeployConfig, StageDeployConfig, merge_pipeline_deploy
+    from vllm_omni.engine.stage_runtime import StageRuntime
+    from vllm_omni.model_executor.models.qwen3_tts.pipeline import QWEN3_TTS_CRISP_HYBRID_PIPELINE
+
+    backend = {
+        "name": "external.crisp.tts.v1",
+        "server_bin": args.server_bin,
+        "server_sha256": args.server_sha256,
+        "talker_file": args.talker_file,
+        "talker_sha256": args.talker_sha256,
+        "codec_file": args.codec_file,
+        "codec_sha256": args.codec_sha256,
+        "punc_file": args.punc_file,
+        "punc_sha256": args.punc_sha256,
+        "log_file": args.server_log,
+        "expected_gpu_name": args.expected_gpu_name,
+        "ggml_vk_visible_devices": "1",
+        "memory_overhead_bytes": 5 << 30,
+        "max_text_bytes": 4096,
+        "max_wav_bytes": 4 << 20,
+        "max_audio_s": 10,
+        "request_timeout_s": 180,
+        "start_timeout_s": 180,
+        "voice": "ryan",
+        "seed": 42,
+    }
+    deploy = DeployConfig(
+        async_chunk=False,
+        stages=[StageDeployConfig(
+            stage_id=0,
+            backend=backend,
+            resource_budget={
+                "capacities": {"host_ram": args.capacity_gib << 30},
+                "demands": {"host_ram": args.reserve_gib << 30},
+            },
+        )],
+    )
+    pipeline = QWEN3_TTS_CRISP_HYBRID_PIPELINE
+    configs = [stage.to_omegaconf() for stage in merge_pipeline_deploy(pipeline, deploy)]
+    runtime = StageRuntime(configs, "local-qwen-tts-crisp", "", stage_init_timeout=180, async_chunk=False)
+    report = {
+        "scope": "complete resident Qwen3-TTS CustomVoice text-to-WAV through Omni StageRuntime/StagePool",
+        "placement": "Radeon Vulkan talker/codec + CPU FP32 code predictor",
+        "warmup_count": args.warmups,
+        "measured_count": args.repeats,
+        "memory_budget": deploy.stages[0].resource_budget,
+    }
+    try:
+        started = time.perf_counter()
+        runtime.initialize()
+        report["startup_s"] = time.perf_counter() - started
+        pool = runtime.stage_pools[0]
+        report["execution_plan"] = pool.stage_client.execution_plan
+        state = SimpleNamespace(sampling_params_list=[None])
+
+        async def request_one(text: str, *, save: bool = False) -> dict:
+            request_id = f"tts-{time.monotonic_ns()}"
+            started = time.perf_counter()
+            await pool.submit_initial(request_id, state, {"text": text, "voice": "ryan", "seed": 42})
+            deadline = time.monotonic() + 180
+            while True:
+                output = pool.poll_graph_output(0)
+                if output is not None:
+                    break
+                if time.monotonic() >= deadline:
+                    raise TimeoutError("Omni CrispASR stage request timed out")
+                await asyncio.sleep(0.002)
+            wall_s = time.perf_counter() - started
+            try:
+                if output.error:
+                    raise RuntimeError(output.error)
+                payload = output.multimodal_output
+                audio, sample_rate = payload["audio"], payload["sr"]
+                metadata = output.custom_output["audio_metadata"]
+                assert sample_rate == 24000 and audio.ndim == 1
+                assert len(audio) == metadata["pcm_frames"] and len(audio) > 0
+                assert float(audio.square().mean().sqrt()) > 1e-5
+                if save:
+                    _save_wav(Path(args.output_wav), audio)
+                return {
+                    "text": text,
+                    "wall_s": wall_s,
+                    "frames": len(audio),
+                    "duration_s": metadata["duration_s"],
+                    "pcm_sha256": metadata["pcm_sha256"],
+                    "stage_event": output.custom_output["stage_event"],
+                }
+            finally:
+                output.release_stage_buffers()
+                await asyncio.sleep(0)
+
+        first = "Hello from the local computer."
+        second = "The blue car is parked beside the library."
+        report["checks"] = [await request_one(first), await request_one(second)]
+        report["warmups"] = [await request_one(first) for _ in range(args.warmups)]
+        report["measured"] = [await request_one(first, save=i == 0) for i in range(args.repeats)]
+        walls = [row["wall_s"] for row in report["measured"]]
+        report["nearest_rank_p50_wall_s"] = _rank(walls, 0.5)
+        report["nearest_rank_p95_wall_s"] = _rank(walls, 0.95)
+        report["all_same_pcm_sha256"] = len({row["pcm_sha256"] for row in report["measured"]}) == 1
+        if args.abort_check:
+            request_id = f"tts-abort-{time.monotonic_ns()}"
+            await pool.submit_initial(request_id, state, {"text": second, "voice": "ryan", "seed": 42})
+            await asyncio.sleep(0.01)
+            await pool.abort_requests([request_id])
+            await asyncio.sleep(0.1)
+            report["abort_check"] = {
+                "stale_output": pool.poll_graph_output(0) is not None,
+                "worker_exited": pool.stage_client._proc.poll() is not None,
+                "ledger_after_abort": runtime.resource_ledger.snapshot(),
+            }
+            if report["abort_check"]["stale_output"] or not report["abort_check"]["worker_exited"]:
+                raise RuntimeError("CrispASR cancellation left stale output or live worker")
+        report["status"] = "passed"
+    except BaseException as exc:
+        report["status"] = "failed"
+        report["error"] = f"{type(exc).__name__}: {exc}"
+        raise
+    finally:
+        runtime.shutdown()
+        if runtime.resource_ledger is not None:
+            report["ledger_after_shutdown"] = runtime.resource_ledger.snapshot()
+        destination = Path(args.output_report)
+        destination.parent.mkdir(parents=True, exist_ok=True)
+        destination.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+        print(json.dumps(report, indent=2))
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/benchmarks/edge_harness/experiments/probe_omni_crisp_tts_entrypoint.py b/benchmarks/edge_harness/experiments/probe_omni_crisp_tts_entrypoint.py
new file mode 100644
index 000000000..80db36b83
--- /dev/null
+++ b/benchmarks/edge_harness/experiments/probe_omni_crisp_tts_entrypoint.py
@@ -0,0 +1,110 @@
+#!/usr/bin/env python3
+"""Verify public AsyncOmni audio delivery from the pinned hybrid TTS stage."""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import hashlib
+import json
+import tempfile
+import time
+from pathlib import Path
+
+import numpy as np
+import yaml
+
+
+async def main() -> None:
+    parser = argparse.ArgumentParser(description=__doc__)
+    for name in (
+        "model_config_dir", "server_bin", "server_sha256", "talker_file", "talker_sha256",
+        "codec_file", "codec_sha256", "punc_file", "punc_sha256", "server_log", "output_report",
+    ):
+        parser.add_argument("--" + name.replace("_", "-"), required=True)
+    args = parser.parse_args()
+
+    from vllm_omni.entrypoints.async_omni import AsyncOmni
+    from vllm_omni.model_executor.models.qwen3_tts.pipeline import QWEN3_TTS_CRISP_HYBRID_PIPELINE
+
+    backend = {
+        "name": "external.crisp.tts.v1",
+        "server_bin": args.server_bin,
+        "server_sha256": args.server_sha256,
+        "talker_file": args.talker_file,
+        "talker_sha256": args.talker_sha256,
+        "codec_file": args.codec_file,
+        "codec_sha256": args.codec_sha256,
+        "punc_file": args.punc_file,
+        "punc_sha256": args.punc_sha256,
+        "log_file": args.server_log,
+        "expected_gpu_name": "AMD Radeon(TM) 890M Graphics",
+        "ggml_vk_visible_devices": "1",
+        "memory_overhead_bytes": 5 << 30,
+        "max_text_bytes": 4096,
+        "max_wav_bytes": 4 << 20,
+        "max_audio_s": 10,
+        "voice": "ryan",
+        "seed": 42,
+    }
+    report = {"entrypoint": "AsyncOmni.generate", "pipeline": QWEN3_TTS_CRISP_HYBRID_PIPELINE.model_type}
+    engine = None
+    try:
+        with tempfile.TemporaryDirectory(prefix="omni-crisp-tts-") as temporary:
+            deployment = Path(temporary) / "deploy.yaml"
+            deployment.write_text(yaml.safe_dump({
+                "pipeline": QWEN3_TTS_CRISP_HYBRID_PIPELINE.model_type,
+                "async_chunk": False,
+                "stages": [{
+                    "stage_id": 0,
+                    "backend": backend,
+                    "resource_budget": {
+                        "capacities": {"host_ram": 16 << 30},
+                        "demands": {"host_ram": 8 << 30},
+                    },
+                }],
+            }), encoding="utf-8")
+            started = time.perf_counter()
+            engine = AsyncOmni(
+                model=str(Path(args.model_config_dir).resolve(strict=True)),
+                deploy_config=str(deployment), stage_init_timeout=180, init_timeout=240,
+            )
+            report["startup_s"] = time.perf_counter() - started
+            outputs = []
+            started = time.perf_counter()
+            async for output in engine.generate(
+                {"text": "Hello from the local computer.", "voice": "ryan", "seed": 42},
+                request_id="public-tts-1",
+            ):
+                if output.error:
+                    raise RuntimeError(output.error)
+                audio = output.multimodal_output["audio"]
+                pcm = np.rint(audio.numpy() * 32768).astype("<i2").tobytes()
+                outputs.append({
+                    "request_id": output.request_id,
+                    "sample_rate": output.multimodal_output["sr"],
+                    "frames": len(audio),
+                    "pcm_sha256": hashlib.sha256(pcm).hexdigest(),
+                    "stage_event": output.custom_output.get("stage_event"),
+                })
+            report["request_wall_s"] = time.perf_counter() - started
+            report["outputs"] = outputs
+            assert len(outputs) == 1 and outputs[0]["request_id"] == "public-tts-1"
+            assert outputs[0]["sample_rate"] == 24000 and outputs[0]["frames"] == 61440
+            assert outputs[0]["pcm_sha256"] == "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539"
+            report["status"] = "passed"
+    except BaseException as exc:
+        report["status"] = "failed"
+        report["error"] = f"{type(exc).__name__}: {exc}"
+        raise
+    finally:
+        if engine is not None:
+            engine.shutdown()
+        destination = Path(args.output_report)
+        destination.parent.mkdir(parents=True, exist_ok=True)
+        destination.write_text(json.dumps(report, indent=2) + "\n", encoding="utf-8")
+        print(json.dumps(report, indent=2))
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md
index e40d8c53f..05e95d1a0 100644
--- a/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/README.md
@@ -1,10 +1,11 @@
 # 2026-09-23 WSL CPU and native Windows expansion
 
-This extends the [60-cell device × model matrix](../e2e_profiling_20260922/evidence/summary/README.md) with previously unrun WSL CPU and native Windows paths. The September 22 timing tables remain historical. MiniCPM-o now has separate 20-request serial text-to-speech timing runs on WSL CPU and native Windows GPU, plus one synthetic image and one synthetic audio input on each platform; each InternVLA timing remains one observation. Spark has a 20-request Omni output-head component profile on the HX370 AMD NPU and separate 20-request **complete text** profiles through standalone llama.cpp on native Windows CPU and Radeon 890M. The latter now also run as bounded whole-session Omni stages, with public AsyncOmni requests checked separately. Qwen3-TTS has separate 20-request complete standalone text-to-WAV profiles on native Windows CPU and a CPU+Radeon 890M hybrid route. None is a sustained or full-modality profile; the TTS standalone paths still need Omni integration. The [checkpoint manifests](../e2e_recovery_20260923/README.md) pin the earlier real weights; the [Spark NPU component record](evidence/spark_amd_npu_output_head/README.md), [standalone CPU/Radeon whole-model record](evidence/spark_radeon890m_llamacpp/README.md), [Omni CPU/Radeon whole-session record](evidence/spark_omni_llamacpp/README.md), [native CPU Qwen TTS record](evidence/qwen_tts_native_cpu/README.md), and [joint CPU+Radeon TTS record](evidence/qwen_tts_radeon890m_joint/README.md) pin their respective artifacts. No model, precision or context was silently substituted.
+This extends the [60-cell device × model matrix](../e2e_profiling_20260922/evidence/summary/README.md) with previously unrun WSL CPU and native Windows paths. The September 22 timing tables remain historical. MiniCPM-o now has separate 20-request serial text-to-speech timing runs on WSL CPU and native Windows GPU, plus one synthetic image and one synthetic audio input on each platform; each InternVLA timing remains one observation. Spark has a 20-request Omni output-head component profile on the HX370 AMD NPU and separate 20-request **complete text** profiles through standalone llama.cpp on native Windows CPU and Radeon 890M. The latter now also run as bounded whole-session Omni stages, with public AsyncOmni requests checked separately. Qwen3-TTS has a 20-request complete standalone text-to-WAV profile on native Windows CPU; the CPU+Radeon 890M hybrid route now has separate standalone and bounded Omni whole-session 20-request profiles. None is a sustained or full-modality profile; the native CPU TTS path still needs Omni integration. The [checkpoint manifests](../e2e_recovery_20260923/README.md) pin the earlier real weights; the [Spark NPU component record](evidence/spark_amd_npu_output_head/README.md), [standalone CPU/Radeon whole-model record](evidence/spark_radeon890m_llamacpp/README.md), [Omni CPU/Radeon whole-session record](evidence/spark_omni_llamacpp/README.md), [native CPU Qwen TTS record](evidence/qwen_tts_native_cpu/README.md), [standalone joint CPU+Radeon TTS record](evidence/qwen_tts_radeon890m_joint/README.md), and [Omni joint TTS record](evidence/qwen_tts_omni_hybrid/README.md) pin their respective artifacts. No model, precision or context was silently substituted.
 
 | Device / model | Verified depth | Observed output and time | Remaining gate |
 |---|---|---|---|
 | Native Windows HX370 CPU+Radeon 890M / Qwen3-TTS 0.6B CustomVoice Q8_0 GGUF plus F16 codec | Scoped complete standalone CrispASR text-to-WAV path with Vulkan talker/codec and CPU FP32 code predictor (P for two named sentences; outside Omni; **not** all-Radeon or NPU) | Two 24 kHz WAVs of 2.56 s and 3.20 s; independently pinned Whisper tiny.en ASR WER 0 on both. One warmup + 20 serial measured resident-server requests for the first prompt had client wall p50/p95 3.310/3.345 s, all PCM hashes identical; RTF 1.29/1.31. The default Vulkan setting fell back to CPU, while all-Vulkan talker+predictor hit a 360-frame runaway cap. | Explicit Omni hybrid stage binding, quality and playback gates, loading-peak admission, cancellation, broader prompts, concurrency and sustained power/thermal behavior; [joint raw evidence](evidence/qwen_tts_radeon890m_joint/README.md), [failed DirectML attempt](evidence/qwen_tts_radeon890m_directml/README.md). |
+| Native Windows HX370 CPU+Radeon 890M / same Qwen3-TTS 0.6B CustomVoice GGUFs | Scoped complete-request Omni CrispASR hybrid stage (P for two named sentences; **not** all-Radeon or NPU) | Both 24 kHz PCM outputs exactly match the earlier standalone hashes. One warmup + 20 serial Omni requests for 2.56 s of audio had 20/20 identical PCM, nearest-rank wall p50/p95 3.298/3.336 s (RTF 1.29/1.30). Public AsyncOmni audio passed; cancellation drained the worker and ledger. | Playable streaming, post-cancel restart, broad speech quality, loading/iGPU peak, concurrency and sustained real-time/thermal gates; [Omni raw evidence](evidence/qwen_tts_omni_hybrid/README.md). |
 | Native Windows HX370 CPU / Qwen3-TTS 0.6B CustomVoice, revision `85e237c12c027371202489a0ec509ded67b5e4b5` | Scoped complete standalone PyTorch text-to-WAV path (P for two named sentences; outside Omni) | Nonzero 24 kHz WAVs of 4.56 s and 3.44 s. For the first prompt, 1 warmup + 20 serial measured complete generations p50/p95 13.94/14.68 s; Whisper tiny.en ASR WER 0.20 on the first and 0.00 on the second sentence. | Omni CPU stage binding and dependency isolation, playable streaming, speech quality, memory loading peak and sustained behavior; [raw evidence](evidence/qwen_tts_native_cpu/README.md). |
 | Native Windows HX370 CPU / Spark-X2.5-1.7B Q4_K_M GGUF, revision `1f7fa33b1245c14730da39e125714ad3a327901b` | Scoped complete standalone llama.cpp text path (P for two named inputs; outside Omni) | Correct `Paris` and 120-item inventory answer `1511`; 1 warmup + 20 uncached serial full inventory requests p50/p95 6.65/6.86 s, 20/20 correct. Stream-close recovery passed. | Omni whole-model stage binding, broader quality, explicit loading-peak admission and sustained behavior; [raw evidence](evidence/spark_radeon890m_llamacpp/README.md). |
 | Native Windows HX370 Radeon 890M / Spark-X2.5-1.7B same GGUF | Scoped complete standalone llama.cpp Vulkan1 text path (P for two named inputs; outside Omni) | All 29 logged layer assignments on Vulkan1, with a separate Vulkan_Host buffer; correct `Paris` and `1511`; 1 warmup + 20 uncached serial inventory requests p50/p95 2.18/2.20 s, 20/20 correct. Stream-close recovery passed. | Omni binding, state and memory contract, broad quality, GPU-memory attribution and sustained power/thermal behavior; [raw evidence](evidence/spark_radeon890m_llamacpp/README.md). |
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md
new file mode 100644
index 000000000..89a1c9bf0
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md
@@ -0,0 +1,28 @@
+# Qwen3-TTS 0.6B CustomVoice on native Windows CPU + Radeon 890M through Omni
+
+This is a **scoped complete-request text-to-audio path** through the registered `qwen3_tts_crisp_hybrid` Omni pipeline and `external.crisp.tts.v1` stage. The stage owns one local CrispASR server process and uses Omni `StageRuntime`/`StagePool` for one-request admission, a declared 8 GiB shared-host-RAM reservation, request/epoch identity, terminal audio event, output acknowledgement and cancellation drain. CrispASR retains the talker, code predictor, codec and their internal state. The stage uses Omni's existing complete-request graph control path; it is not an ONNX graph and does not yet provide playable incremental audio or restart after an in-flight cancellation.
+
+The host was Windows 11 build 26200 on Ryzen AI 9 HX 370, Radeon 890M driver `32.0.22018.6001`, with an RTX 5090 Laptop GPU also installed. The active Windows power scheme was observed as `Performance` after the run; temperature and package power were not sampled. The selected hybrid route exposed only the Radeon through `GGML_VK_VISIBLE_DEVICES=1`, used `CRISPASR_QWEN3_TTS_VULKAN_NATIVE=1` for the talker/codec, and pinned the five-layer code predictor to CPU FP32 with `CRISPASR_QWEN3_TTS_CP_BACKEND=cpu-f32`. The [server log](omni_crisp_tts_pinned_profile20_20260923.log) identifies the Radeon as Vulkan0, the CPU-pinned predictor and Vulkan0 codec loading. This is route/log evidence, not per-operator GPU profiling. The present backend is intentionally restricted to this verified HX370/890M shared-RAM route.
+
+Artifacts checked by SHA-256 **before launch**:
+
+| Artifact | Revision / build | SHA-256 |
+|---|---|---|
+| CrispASR `crispasr.exe` | v0.8.35, git `00c9e752`; CPU/Vulkan build | `999c6af252077195adb723b96da9ee5d3cd291dc7c7a4b32dc493ae148035529` |
+| CustomVoice Q8_0 talker GGUF | `cstr/qwen3-tts-0.6b-customvoice-GGUF` revision `29094dcb46f19b0898616aa9aeb876aa9aa6e086` | `5227dcbc4df7c5533341d111cc469fa491a48e722b23dd10f553181b52dff2d9` |
+| F16 codec GGUF | `cstr/qwen3-tts-tokenizer-12hz-GGUF` revision `344e279551177654762e5c7d39e217834325f222` | `70dc95dbfdd9aa5d9d406236ff771d061bf17b0cda02a72513953355606e719b` |
+| Auto-loaded punctuation GGUF, now passed explicitly | local cached `fireredpunc-q4_k.gguf`; upstream revision not recorded | `faf4a43e3135bc307a66194685af00f756e6f4c28c7d9e2dd8f3517cddca5c45` |
+
+The weights and executable remain in local model storage, not Git. The talker and codec revisions are pinned in the [earlier standalone record](../qwen_tts_radeon890m_joint/README.md). The converter's source-checkpoint revision, calibration and full same-checkpoint numerical parity remain unestablished; this is not a native HF-safetensors precision result. The punctuation model was automatically loaded by the earlier resident server; this integration explicitly hashes and passes the exact cached file to remove an unpinned startup dependency.
+
+| Named input / profile | Omni result |
+|---|---|
+| “Hello from the local computer.”, Ryan, seed 42 | 61,440 samples, 2.56 s at 24 kHz mono; PCM SHA-256 `57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539` |
+| “The blue car is parked beside the library.”, Ryan, seed 42 | 76,800 samples, 3.20 s at 24 kHz mono; PCM SHA-256 `4832877787577d2e47b51f325a09e17df37b7f2a95ed5297b37259977286476d` |
+| First sentence, one excluded warmup plus 20 serial complete requests | 20/20 outputs had the identical PCM SHA-256. Nearest-rank Omni request-wall p50/p95 **3.298/3.336 s**, or RTF **1.29/1.30** against 2.56 s of audio. Stage startup, including artifact hashing and load, was 4.64 s. |
+
+Both named PCM hashes match the [earlier standalone output](../qwen_tts_radeon890m_joint/README.md) exactly. The [raw profile JSON](omni_crisp_tts_pinned_profile20_20260923.json) retains all request times, frame counts, stage events, placement plan and ledger snapshots. The [first measured WAV](omni_crisp_tts_pinned_profile20_20260923.wav) was independently transcribed as the exact sentence by pinned Whisper tiny.en, WER 0 in the [ASR report](omni_crisp_tts_pinned_profile20_asr_20260923.json). ASR agreement is only an intelligibility proxy, not speaker or perceptual quality. The [public AsyncOmni JSON](omni_crisp_tts_pinned_public_20260923.json) records one terminal audio output of the same PCM hash using a real Qwen3-TTS HF config directory for pipeline resolution; the GGUF weights, rather than the HF weights, executed.
+
+The profile cancelled a live request after the measured series, observed no stale audio, terminated its owned process and released the ledger with no quarantine. It does not prove the stage can continue serving after that active cancellation. A [4 GiB reservation was refused](omni_crisp_tts_pinned_reject_memory_20260923.json) before load, and a [different declared GPU was refused](omni_crisp_tts_pinned_reject_device_20260923.json); both left the ledger empty. The admitted 8 GiB includes 1.38 GB of pinned GGUF files plus a declared 5 GiB allowance for KV, activations, workspace, transfers, loading and headroom; loaded process RSS was 2,723,655,680 bytes. Neither number is a measured loading peak or complete iGPU-memory attribution. The route remained slower than real-time generation and was not checked for continuous playback, concurrent requests, longer speech, thermal throttling or 30-minute sustained behavior, so M2 release qualification remains open. The CPU+iGPU+NPU joint row is unchanged because no NPU participated.
+
+Reproduce with the [stage profile script](../../../../experiments/probe_omni_crisp_tts.py) and [public API script](../../../../experiments/probe_omni_crisp_tts_entrypoint.py) on native Windows Python/vLLM 0.29 with `PYTHONPATH` pointing to this checkout, `PYTHONUTF8=1`, and `VLLM_ENABLE_V1_MULTIPROCESSING=0`. Supply local paths and the four hashes above, plus `--server-log`, `--output-report`, and for the stage profile `--output-wav --warmups 1 --repeats 20 --abort-check`. The public probe also accepts `--model-config-dir`. No mobile or embedded device is implied by this Windows run.
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.driver.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.driver.log
new file mode 100644
index 000000000..a132623da
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.driver.log
@@ -0,0 +1,495 @@
+INFO 09-23 20:44:07 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
+INFO 09-23 20:44:08 [patch.py:510] inductor factorable-divisibility patch: installed.
+INFO 09-23 20:44:11 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
+WARNING 09-23 20:44:11 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 13] Permission denied: 'C:\\Users\\zhout\\w2'
+{
+  "scope": "complete resident Qwen3-TTS CustomVoice text-to-WAV through Omni StageRuntime/StagePool",
+  "placement": "Radeon Vulkan talker/codec + CPU FP32 code predictor",
+  "warmup_count": 1,
+  "measured_count": 20,
+  "memory_budget": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "demands": {
+      "host_ram": 8589934592
+    }
+  },
+  "startup_s": 4.644805799936876,
+  "execution_plan": {
+    "backend": "external.crisp.tts.v1",
+    "stage_id": 0,
+    "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+    "worker_pid": 25068,
+    "artifact_sha256": {
+      "server_sha256": "999c6af252077195adb723b96da9ee5d3cd291dc7c7a4b32dc493ae148035529",
+      "talker_sha256": "5227dcbc4df7c5533341d111cc469fa491a48e722b23dd10f553181b52dff2d9",
+      "codec_sha256": "70dc95dbfdd9aa5d9d406236ff771d061bf17b0cda02a72513953355606e719b",
+      "punc_sha256": "faf4a43e3135bc307a66194685af00f756e6f4c28c7d9e2dd8f3517cddca5c45"
+    },
+    "placement": "Radeon Vulkan0 talker/codec + CPU FP32 code predictor",
+    "expected_gpu_name": "AMD Radeon(TM) 890M Graphics",
+    "ggml_vk_visible_devices": "1",
+    "loaded_rss_bytes": 2723655680,
+    "reserved_bytes": {
+      "host_ram": 8589934592
+    },
+    "memory_overhead_bytes": 5368709120,
+    "max_text_bytes": 4096,
+    "max_wav_bytes": 4194304,
+    "max_audio_s": 10.0,
+    "voice": "ryan",
+    "seed": 42,
+    "request_capacity": 1,
+    "stateful_session": "CrispASR-owned; complete WAV per request",
+    "evidence": "B"
+  },
+  "checks": [
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.339751899940893,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622190359000000",
+        "stage_id": 0,
+        "epoch": 1,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "The blue car is parked beside the library.",
+      "wall_s": 4.12443029996939,
+      "frames": 76800,
+      "duration_s": 3.2,
+      "pcm_sha256": "4832877787577d2e47b51f325a09e17df37b7f2a95ed5297b37259977286476d",
+      "stage_event": {
+        "request_id": "tts-622193703000000",
+        "stage_id": 0,
+        "epoch": 2,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    }
+  ],
+  "warmups": [
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.28562800004147,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622197828000000",
+        "stage_id": 0,
+        "epoch": 3,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    }
+  ],
+  "measured": [
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3386760000139475,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622201109000000",
+        "stage_id": 0,
+        "epoch": 4,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3058673000195995,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622204453000000",
+        "stage_id": 0,
+        "epoch": 5,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.304257000098005,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622207750000000",
+        "stage_id": 0,
+        "epoch": 6,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3195733999600634,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622211062000000",
+        "stage_id": 0,
+        "epoch": 7,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.275992800015956,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622214375000000",
+        "stage_id": 0,
+        "epoch": 8,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.262796200113371,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622217656000000",
+        "stage_id": 0,
+        "epoch": 9,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.237368700094521,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622220921000000",
+        "stage_id": 0,
+        "epoch": 10,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.2933594998903573,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622224156000000",
+        "stage_id": 0,
+        "epoch": 11,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.301911799935624,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622227453000000",
+        "stage_id": 0,
+        "epoch": 12,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3207750000292435,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622230750000000",
+        "stage_id": 0,
+        "epoch": 13,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3186378000536934,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622234078000000",
+        "stage_id": 0,
+        "epoch": 14,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3360501000424847,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622237390000000",
+        "stage_id": 0,
+        "epoch": 15,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.287759199971333,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622240718000000",
+        "stage_id": 0,
+        "epoch": 16,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.2255696000065655,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622244015000000",
+        "stage_id": 0,
+        "epoch": 17,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.2808044999837875,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622247234000000",
+        "stage_id": 0,
+        "epoch": 18,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.224921299959533,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622250515000000",
+        "stage_id": 0,
+        "epoch": 19,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3109662000788376,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622253750000000",
+        "stage_id": 0,
+        "epoch": 20,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.270013100001961,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622257062000000",
+        "stage_id": 0,
+        "epoch": 21,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.333422400057316,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622260328000000",
+        "stage_id": 0,
+        "epoch": 22,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.2983070999616757,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622263656000000",
+        "stage_id": 0,
+        "epoch": 23,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    }
+  ],
+  "nearest_rank_p50_wall_s": 3.2983070999616757,
+  "nearest_rank_p95_wall_s": 3.3360501000424847,
+  "all_same_pcm_sha256": true,
+  "abort_check": {
+    "stale_output": false,
+    "worker_exited": true,
+    "ledger_after_abort": {
+      "capacities": {
+        "host_ram": 17179869184
+      },
+      "reserved": {
+        "host_ram": 0
+      },
+      "owners": [],
+      "quarantined": []
+    }
+  },
+  "status": "passed",
+  "ledger_after_shutdown": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "reserved": {
+      "host_ram": 0
+    },
+    "owners": [],
+    "quarantined": []
+  }
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.json
new file mode 100644
index 000000000..16d1e3f2d
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.json
@@ -0,0 +1,491 @@
+{
+  "scope": "complete resident Qwen3-TTS CustomVoice text-to-WAV through Omni StageRuntime/StagePool",
+  "placement": "Radeon Vulkan talker/codec + CPU FP32 code predictor",
+  "warmup_count": 1,
+  "measured_count": 20,
+  "memory_budget": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "demands": {
+      "host_ram": 8589934592
+    }
+  },
+  "startup_s": 4.644805799936876,
+  "execution_plan": {
+    "backend": "external.crisp.tts.v1",
+    "stage_id": 0,
+    "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+    "worker_pid": 25068,
+    "artifact_sha256": {
+      "server_sha256": "999c6af252077195adb723b96da9ee5d3cd291dc7c7a4b32dc493ae148035529",
+      "talker_sha256": "5227dcbc4df7c5533341d111cc469fa491a48e722b23dd10f553181b52dff2d9",
+      "codec_sha256": "70dc95dbfdd9aa5d9d406236ff771d061bf17b0cda02a72513953355606e719b",
+      "punc_sha256": "faf4a43e3135bc307a66194685af00f756e6f4c28c7d9e2dd8f3517cddca5c45"
+    },
+    "placement": "Radeon Vulkan0 talker/codec + CPU FP32 code predictor",
+    "expected_gpu_name": "AMD Radeon(TM) 890M Graphics",
+    "ggml_vk_visible_devices": "1",
+    "loaded_rss_bytes": 2723655680,
+    "reserved_bytes": {
+      "host_ram": 8589934592
+    },
+    "memory_overhead_bytes": 5368709120,
+    "max_text_bytes": 4096,
+    "max_wav_bytes": 4194304,
+    "max_audio_s": 10.0,
+    "voice": "ryan",
+    "seed": 42,
+    "request_capacity": 1,
+    "stateful_session": "CrispASR-owned; complete WAV per request",
+    "evidence": "B"
+  },
+  "checks": [
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.339751899940893,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622190359000000",
+        "stage_id": 0,
+        "epoch": 1,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "The blue car is parked beside the library.",
+      "wall_s": 4.12443029996939,
+      "frames": 76800,
+      "duration_s": 3.2,
+      "pcm_sha256": "4832877787577d2e47b51f325a09e17df37b7f2a95ed5297b37259977286476d",
+      "stage_event": {
+        "request_id": "tts-622193703000000",
+        "stage_id": 0,
+        "epoch": 2,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    }
+  ],
+  "warmups": [
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.28562800004147,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622197828000000",
+        "stage_id": 0,
+        "epoch": 3,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    }
+  ],
+  "measured": [
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3386760000139475,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622201109000000",
+        "stage_id": 0,
+        "epoch": 4,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3058673000195995,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622204453000000",
+        "stage_id": 0,
+        "epoch": 5,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.304257000098005,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622207750000000",
+        "stage_id": 0,
+        "epoch": 6,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3195733999600634,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622211062000000",
+        "stage_id": 0,
+        "epoch": 7,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.275992800015956,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622214375000000",
+        "stage_id": 0,
+        "epoch": 8,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.262796200113371,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622217656000000",
+        "stage_id": 0,
+        "epoch": 9,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.237368700094521,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622220921000000",
+        "stage_id": 0,
+        "epoch": 10,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.2933594998903573,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622224156000000",
+        "stage_id": 0,
+        "epoch": 11,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.301911799935624,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622227453000000",
+        "stage_id": 0,
+        "epoch": 12,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3207750000292435,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622230750000000",
+        "stage_id": 0,
+        "epoch": 13,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3186378000536934,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622234078000000",
+        "stage_id": 0,
+        "epoch": 14,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3360501000424847,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622237390000000",
+        "stage_id": 0,
+        "epoch": 15,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.287759199971333,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622240718000000",
+        "stage_id": 0,
+        "epoch": 16,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.2255696000065655,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622244015000000",
+        "stage_id": 0,
+        "epoch": 17,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.2808044999837875,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622247234000000",
+        "stage_id": 0,
+        "epoch": 18,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.224921299959533,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622250515000000",
+        "stage_id": 0,
+        "epoch": 19,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.3109662000788376,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622253750000000",
+        "stage_id": 0,
+        "epoch": 20,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.270013100001961,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622257062000000",
+        "stage_id": 0,
+        "epoch": 21,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.333422400057316,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622260328000000",
+        "stage_id": 0,
+        "epoch": 22,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    },
+    {
+      "text": "Hello from the local computer.",
+      "wall_s": 3.2983070999616757,
+      "frames": 61440,
+      "duration_s": 2.56,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "tts-622263656000000",
+        "stage_id": 0,
+        "epoch": 23,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "0ee68a7641724a1b9bba5eadc021a3fb",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    }
+  ],
+  "nearest_rank_p50_wall_s": 3.2983070999616757,
+  "nearest_rank_p95_wall_s": 3.3360501000424847,
+  "all_same_pcm_sha256": true,
+  "abort_check": {
+    "stale_output": false,
+    "worker_exited": true,
+    "ledger_after_abort": {
+      "capacities": {
+        "host_ram": 17179869184
+      },
+      "reserved": {
+        "host_ram": 0
+      },
+      "owners": [],
+      "quarantined": []
+    }
+  },
+  "status": "passed",
+  "ledger_after_shutdown": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "reserved": {
+      "host_ram": 0
+    },
+    "owners": [],
+    "quarantined": []
+  }
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.log
new file mode 100644
index 000000000..fec344e9f
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_20260923.log
@@ -0,0 +1,167 @@
+crispasr 0.8.35 (git 00c9e752, Release) [backends: cpu,vulkan]
+=== build info ===
+  version       : 0.8.35
+  git sha       : 00c9e752
+  git date      : 2026-09-21T20:44:31Z
+  git subject   : release-_bump_VERSION_to_0.8.35
+  build date    : 2026-09-21T21:04:24Z
+  build type    : Release
+  compiler      : msvc 1951
+  os            : windows
+  arch          : x86_64
+  ggml backends : cpu,vulkan
+
+=== runtime env ===
+  PATH                         = C:\Users\zhout\w2\omni029venv\Lib\site-packages\cv2\../../x64/vc17/bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\powershell;C:\Users\zhout\.codex\tmp\arg0\codex-arg05Acw4D;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\libheif\libheif\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\jxrlib\jxrlib\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\poppler\Library\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\bin\override;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin\x64;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin;C:\Windows\system32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0\;C:\Windows\System32\OpenSSH\;C:\Program Files (x86)\NVIDIA Corporation\PhysX\Common;C:\Program Files\NVIDIA Corporation\NVIDIA App\NvDLISR;C:\Windows\System32\WindowsPowerShell\v1.0;C:\Program Files\Git\cmd;C:\Program Files\NVIDIA Corporation\Nsight Compute 2026.3.0\;C:\Users\zhout\.cargo\bin;C:\Users\zhout\AppData\Local\Programs\Python\Python312\Scripts\;C:\Users\zhout\AppData\Local\Programs\Python\Python312\;C:\Users\zhout\AppData\Local\Programs\Python\Launcher\;c:\Users\zhout\AppData\Local\Programs\cursor\resources\app\codeBin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\powershell;C:\Users\zhout\AppData\Local\Microsoft\WindowsApps;C:\Users\zhout\AppData\Local\Programs\cursor\resources\app\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\bin\fallback;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\node\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\git\cmd;C:\Program Files\WindowsApps\OpenAI.Codex_26.917.6896.0_x64__2p2nqsd0c76g0\app\resources
+  LD_LIBRARY_PATH              = (unset)
+  LIBRARY_PATH                 = (unset)
+  CUDA_HOME                    = (unset)
+  CUDA_PATH                    = C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4
+  CUDA_VISIBLE_DEVICES         = 1
+  NVIDIA_VISIBLE_DEVICES       = (unset)
+  NVIDIA_DRIVER_CAPABILITIES   = (unset)
+  GGML_CUDA_DEBUG              = 1
+  GGML_VK_VISIBLE_DEVICES      = 1
+  GGML_VK_PIPELINE_CACHE_DEBUG = 1
+  CRISPASR_USE_CUDA_COMPAT     = (unset)
+  CRISPASR_BACKEND             = (unset)
+  CRISPASR_CACHE_DIR           = (unset)
+
+=== ggml backends + devices ===
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon(TM) 890M Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 1 | fp4: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+  registered backends: 2
+    [0] Vulkan (devices: 1)
+    [1] CPU (devices: 1)
+  registered devices : 2
+    [0] igpu   name=Vulkan0 desc=AMD Radeon(TM) 890M Graphics mem=46080/48505 MiB id=?
+    [1] cpu    name=CPU desc=AMD Ryzen AI 9 HX 370 w/ Radeon 890M            mem=30446/64633 MiB id=?
+
+cpu-isa: OK — built for AVX2/FMA/F16C/BMI2/AVX; host CPU supports the build's ISA baseline
+crispasr_init_gpu_backend: using preferred GPU backend: Vulkan0
+qwen3_tts: code_pred CPU-pinned (72 tensors, 432 MB, copy=f32)
+qwen3_tts: kv cache 448 MiB (head_dim=128 max_ctx=4096 n_kv=8 n_layers=28)
+qwen3_tts: loaded C:\Users\zhout\w2\models\crisp_qwen3tts_0.6b\qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf  (talker 28L/1024  code_pred 5L  vocab 3072)
+qwen3_tts: codec: GPU default - loading weights onto Vulkan0
+qwen3_tts: codec loaded from 'C:\Users\zhout\w2\models\crisp_qwen3tts_0.6b\qwen3-tts-tokenizer-12hz.gguf'  (8L d=512/1024  rvq=16)
+qwen3_tts: codec encoder loaded (SEANet+xfmr+RVQ)
+crispasr[qwen3-tts]: codec loaded from 'C:\Users\zhout\w2\models\crisp_qwen3tts_0.6b\qwen3-tts-tokenizer-12hz.gguf'
+crispasr-server: warmup completed in 0 ms
+crispasr-server: backend 'qwen3-tts-customvoice' loaded, model 'C:\Users\zhout\w2\models\crisp_qwen3tts_0.6b\qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf'
+fireredpunc: 12L, d=768, ffn=3072, heads=12, vocab=21128, labels=5
+crispasr_init_gpu_backend: using preferred GPU backend: Vulkan0
+crispasr-server: loaded punctuation model 'C:\Users\zhout\.cache\crispasr\fireredpunc-q4_k.gguf'
+
+crispasr-server: listening on 127.0.0.1:51785
+  POST /inference                  — upload audio (native JSON)
+  POST /v1/audio/transcriptions    — OpenAI-compatible API
+  POST /v1/audio/speech            — TTS (OpenAI-compatible)
+  POST /load                       — hot-swap model
+  GET  /health                     — server status
+  GET  /backends                   — list backends
+  GET  /v1/models                  — model info
+  GET  /v1/voices                  — list voices in --voice-dir
+       /voices, /v1/audio/voices   — aliases (llama-swap compat)
+  POST /v1/voices                  — upload voice file (multipart)
+  DELETE /v1/voices/:name          — delete voice file
+crispasr-server: warning: --voice-dir not set; /v1/voices will return empty and /v1/audio/speech will reject requests with a 'voice' field
+
+crispasr: warning: 'qwen3-tts-customvoice' does not record whether its preset voice belongs to a real person, so no spoken AI
+  disclosure was added. If the voice is that of an identifiable individual, the output is a
+  deep fake under EU AI Act Art. 3(60) and you carry the Art. 50(4) duty to disclose it.
+  Pass --speaker-identity real_person to have CrispASR prepend the disclosure, or
+  --speaker-identity synthetic to silence this.
+qwen3_tts: selected CustomVoice speaker 'ryan' (spk_id=3061)
+crispasr[qwen3-tts]: CustomVoice speaker = 'ryan' (available: aiden, dylan, eric, ono_anna, ryan, serena, sohee, uncle_fu, vivian)
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3059 ms  codec=210 ms  total=3269 ms  audio=2.6 s  rtf=1.277
+crispasr-server: synthesized 2.6s audio in 3.28s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=10 + final=1 = T=19 trailing=1
+qwen3_tts: produced 40 frames × 16 codebooks = 640 codes
+qwen3_tts: perf — codes=3822 ms  codec=258 ms  total=4080 ms  audio=3.2 s  rtf=1.275
+crispasr-server: synthesized 3.2s audio in 4.11s (RTF=1.29) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3053 ms  codec=210 ms  total=3263 ms  audio=2.6 s  rtf=1.275
+crispasr-server: synthesized 2.6s audio in 3.26s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3103 ms  codec=207 ms  total=3309 ms  audio=2.6 s  rtf=1.293
+crispasr-server: synthesized 2.6s audio in 3.31s (RTF=1.29) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3067 ms  codec=206 ms  total=3272 ms  audio=2.6 s  rtf=1.278
+crispasr-server: synthesized 2.6s audio in 3.27s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3063 ms  codec=205 ms  total=3269 ms  audio=2.6 s  rtf=1.277
+crispasr-server: synthesized 2.6s audio in 3.27s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3073 ms  codec=214 ms  total=3287 ms  audio=2.6 s  rtf=1.284
+crispasr-server: synthesized 2.6s audio in 3.29s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3037 ms  codec=206 ms  total=3243 ms  audio=2.6 s  rtf=1.267
+crispasr-server: synthesized 2.6s audio in 3.24s (RTF=1.27) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3034 ms  codec=204 ms  total=3237 ms  audio=2.6 s  rtf=1.265
+crispasr-server: synthesized 2.6s audio in 3.24s (RTF=1.27) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=2998 ms  codec=207 ms  total=3205 ms  audio=2.6 s  rtf=1.252
+crispasr-server: synthesized 2.6s audio in 3.21s (RTF=1.25) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3062 ms  codec=204 ms  total=3266 ms  audio=2.6 s  rtf=1.276
+crispasr-server: synthesized 2.6s audio in 3.27s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3064 ms  codec=206 ms  total=3270 ms  audio=2.6 s  rtf=1.277
+crispasr-server: synthesized 2.6s audio in 3.27s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3091 ms  codec=205 ms  total=3296 ms  audio=2.6 s  rtf=1.288
+crispasr-server: synthesized 2.6s audio in 3.30s (RTF=1.29) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3099 ms  codec=209 ms  total=3308 ms  audio=2.6 s  rtf=1.292
+crispasr-server: synthesized 2.6s audio in 3.31s (RTF=1.29) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3100 ms  codec=204 ms  total=3305 ms  audio=2.6 s  rtf=1.291
+crispasr-server: synthesized 2.6s audio in 3.31s (RTF=1.29) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3069 ms  codec=208 ms  total=3277 ms  audio=2.6 s  rtf=1.280
+crispasr-server: synthesized 2.6s audio in 3.28s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=2983 ms  codec=206 ms  total=3189 ms  audio=2.6 s  rtf=1.246
+crispasr-server: synthesized 2.6s audio in 3.19s (RTF=1.25) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3065 ms  codec=205 ms  total=3270 ms  audio=2.6 s  rtf=1.277
+crispasr-server: synthesized 2.6s audio in 3.27s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=2988 ms  codec=204 ms  total=3192 ms  audio=2.6 s  rtf=1.247
+crispasr-server: synthesized 2.6s audio in 3.19s (RTF=1.25) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3072 ms  codec=206 ms  total=3278 ms  audio=2.6 s  rtf=1.280
+crispasr-server: synthesized 2.6s audio in 3.28s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3054 ms  codec=205 ms  total=3259 ms  audio=2.6 s  rtf=1.273
+crispasr-server: synthesized 2.6s audio in 3.26s (RTF=1.27) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3093 ms  codec=206 ms  total=3299 ms  audio=2.6 s  rtf=1.289
+crispasr-server: synthesized 2.6s audio in 3.30s (RTF=1.29) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3056 ms  codec=208 ms  total=3263 ms  audio=2.6 s  rtf=1.275
+crispasr-server: synthesized 2.6s audio in 3.27s (RTF=1.28) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_asr_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_asr_20260923.json
new file mode 100644
index 000000000..e4f2803d6
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_asr_20260923.json
@@ -0,0 +1,16 @@
+{
+  "scope": "local Whisper tiny.en ASR proxy for generated speech; no perceptual quality claim",
+  "asr_model": "openai/whisper-tiny.en",
+  "asr_revision": "87c7102498dcde7456f24cfd30239ca606ed9063",
+  "asr_weights_sha256": "db59695928ded6043adaef491a53ef4e12da9611184d77c53baa691a60b958ad",
+  "transformers": "4.57.3",
+  "torch": "2.4.1+cpu",
+  "wav_sha256": "1bb9590f367ab531c66fc20d9b8aed9f0477fde67f2bdbe16e60de5a40b17e27",
+  "input_sample_rate": 24000,
+  "input_duration_s": 2.56,
+  "reference_text": "Hello from the local computer.",
+  "transcript": "Hello from the local computer.",
+  "word_error_rate": 0.0,
+  "asr_load_s": 1.1599759999662638,
+  "asr_inference_s": 1.4488219999475405
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_asr_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_asr_20260923.log
new file mode 100644
index 000000000..ff4970343
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_profile20_asr_20260923.log
@@ -0,0 +1,17 @@
+Using custom `forced_decoder_ids` from the (generation) config. This is deprecated in favor of the `task` and `language` flags/config options.
+{
+  "scope": "local Whisper tiny.en ASR proxy for generated speech; no perceptual quality claim",
+  "asr_model": "openai/whisper-tiny.en",
+  "asr_revision": "87c7102498dcde7456f24cfd30239ca606ed9063",
+  "asr_weights_sha256": "db59695928ded6043adaef491a53ef4e12da9611184d77c53baa691a60b958ad",
+  "transformers": "4.57.3",
+  "torch": "2.4.1+cpu",
+  "wav_sha256": "1bb9590f367ab531c66fc20d9b8aed9f0477fde67f2bdbe16e60de5a40b17e27",
+  "input_sample_rate": 24000,
+  "input_duration_s": 2.56,
+  "reference_text": "Hello from the local computer.",
+  "transcript": "Hello from the local computer.",
+  "word_error_rate": 0.0,
+  "asr_load_s": 1.1599759999662638,
+  "asr_inference_s": 1.4488219999475405
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.driver.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.driver.log
new file mode 100644
index 000000000..0bab5586c
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.driver.log
@@ -0,0 +1,52 @@
+INFO 09-23 20:46:11 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
+INFO 09-23 20:46:11 [patch.py:510] inductor factorable-divisibility patch: installed.
+INFO 09-23 20:46:15 [omni_base.py:216] [AsyncOmni] Initializing with model C:\Users\zhout\w2\models\Qwen3-TTS-0.6B-85e237c
+INFO 09-23 20:46:15 [async_omni_engine.py:172] [AsyncOmniEngine] Initializing with model C:\Users\zhout\w2\models\Qwen3-TTS-0.6B-85e237c
+ERROR 09-23 20:46:15 [repo_utils.py:119] Error parsing config for C:\Users\zhout\w2\models\Qwen3-TTS-0.6B-85e237c: The checkpoint you are trying to load has model type `qwen3_tts` but Transformers does not recognize this architecture. This could be because of an issue with the checkpoint, or because your version of Transformers is out of date.
+
+ERROR 09-23 20:46:15 [repo_utils.py:119] 
+
+ERROR 09-23 20:46:15 [repo_utils.py:119] You can update Transformers with the command `pip install --upgrade transformers`. If this does not work, and the checkpoint is very new, then there may not be a release version that supports this model yet. In this case, you can get the most up-to-date code by installing Transformers from source with the command `pip install git+https://github.com/huggingface/transformers.git`, retrying 1 of 2
+ERROR 09-23 20:46:17 [repo_utils.py:117] Error parsing config for C:\Users\zhout\w2\models\Qwen3-TTS-0.6B-85e237c: The checkpoint you are trying to load has model type `qwen3_tts` but Transformers does not recognize this architecture. This could be because of an issue with the checkpoint, or because your version of Transformers is out of date.
+
+ERROR 09-23 20:46:17 [repo_utils.py:117] 
+
+ERROR 09-23 20:46:17 [repo_utils.py:117] You can update Transformers with the command `pip install --upgrade transformers`. If this does not work, and the checkpoint is very new, then there may not be a release version that supports this model yet. In this case, you can get the most up-to-date code by installing Transformers from source with the command `pip install git+https://github.com/huggingface/transformers.git`
+INFO 09-23 20:46:17 [async_omni_engine.py:290] [AsyncOmniEngine] Launching Orchestrator thread with 1 stages
+INFO 09-23 20:46:17 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
+INFO 09-23 20:46:17 [initialization.py:356] Loaded OmniTransferConfig with 0 connector configurations
+INFO 09-23 20:46:21 [orchestrator.py:604] [Orchestrator] Starting event loop
+INFO 09-23 20:46:21 [async_omni_engine.py:323] [AsyncOmniEngine] Orchestrator ready with 1 stages
+INFO 09-23 20:46:21 [omni_base.py:235] [AsyncOmni] AsyncOmniEngine initialized in 6.84 seconds
+INFO 09-23 20:46:21 [omni_base.py:255] [AsyncOmni] Initialized with 1 stages for model C:\Users\zhout\w2\models\Qwen3-TTS-0.6B-85e237c
+INFO 09-23 20:46:25 [omni_base.py:781] [AsyncOmni] Shutting down
+INFO 09-23 20:46:25 [async_omni_engine.py:1807] [AsyncOmniEngine] Shutting down Orchestrator
+INFO 09-23 20:46:25 [orchestrator.py:705] [Orchestrator] Received shutdown signal
+INFO 09-23 20:46:25 [orchestrator.py:3092] [Orchestrator] Shutting down all 1 client(s)
+INFO 09-23 20:46:25 [stage_pool.py:1354] [StagePool] Stage 0 replica 0 shut down
+{
+  "entrypoint": "AsyncOmni.generate",
+  "pipeline": "qwen3_tts_crisp_hybrid",
+  "startup_s": 6.844502300024033,
+  "request_wall_s": 3.4202576000243425,
+  "outputs": [
+    {
+      "request_id": "public-tts-1",
+      "sample_rate": 24000,
+      "frames": 61440,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "public-tts-1-b42cf9e7",
+        "stage_id": 0,
+        "epoch": 1,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "533c35d38e094bdf95f88ccac2dbab01",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    }
+  ],
+  "status": "passed"
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.json
new file mode 100644
index 000000000..b8afb820d
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.json
@@ -0,0 +1,26 @@
+{
+  "entrypoint": "AsyncOmni.generate",
+  "pipeline": "qwen3_tts_crisp_hybrid",
+  "startup_s": 6.844502300024033,
+  "request_wall_s": 3.4202576000243425,
+  "outputs": [
+    {
+      "request_id": "public-tts-1",
+      "sample_rate": 24000,
+      "frames": 61440,
+      "pcm_sha256": "57f60228e2b456f6dfefc7e49df862c7bff6ba7cf39231ef1824f615948a1539",
+      "stage_event": {
+        "request_id": "public-tts-1-b42cf9e7",
+        "stage_id": 0,
+        "epoch": 1,
+        "seq": 1,
+        "kind": "audio",
+        "worker_generation": "533c35d38e094bdf95f88ccac2dbab01",
+        "buffers": [],
+        "terminal": true,
+        "error": null
+      }
+    }
+  ],
+  "status": "passed"
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.log
new file mode 100644
index 000000000..f922ba503
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_public_20260923.log
@@ -0,0 +1,79 @@
+crispasr 0.8.35 (git 00c9e752, Release) [backends: cpu,vulkan]
+=== build info ===
+  version       : 0.8.35
+  git sha       : 00c9e752
+  git date      : 2026-09-21T20:44:31Z
+  git subject   : release-_bump_VERSION_to_0.8.35
+  build date    : 2026-09-21T21:04:24Z
+  build type    : Release
+  compiler      : msvc 1951
+  os            : windows
+  arch          : x86_64
+  ggml backends : cpu,vulkan
+
+=== runtime env ===
+  PATH                         = C:\Users\zhout\w2\omni029venv\Lib\site-packages\cv2\../../x64/vc17/bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\powershell;C:\Users\zhout\.codex\tmp\arg0\codex-arg05Acw4D;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\libheif\libheif\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\jxrlib\jxrlib\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\poppler\Library\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\bin\override;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin\x64;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4\bin;C:\Windows\system32;C:\Windows;C:\Windows\System32\Wbem;C:\Windows\System32\WindowsPowerShell\v1.0\;C:\Windows\System32\OpenSSH\;C:\Program Files (x86)\NVIDIA Corporation\PhysX\Common;C:\Program Files\NVIDIA Corporation\NVIDIA App\NvDLISR;C:\Windows\System32\WindowsPowerShell\v1.0;C:\Program Files\Git\cmd;C:\Program Files\NVIDIA Corporation\Nsight Compute 2026.3.0\;C:\Users\zhout\.cargo\bin;C:\Users\zhout\AppData\Local\Programs\Python\Python312\Scripts\;C:\Users\zhout\AppData\Local\Programs\Python\Python312\;C:\Users\zhout\AppData\Local\Programs\Python\Launcher\;c:\Users\zhout\AppData\Local\Programs\cursor\resources\app\codeBin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\powershell;C:\Users\zhout\AppData\Local\Microsoft\WindowsApps;C:\Users\zhout\AppData\Local\Programs\cursor\resources\app\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\bin\fallback;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\node\bin;C:\Users\zhout\.cache\codex-runtimes\codex-primary-runtime\dependencies\native\git\cmd;C:\Program Files\WindowsApps\OpenAI.Codex_26.917.6896.0_x64__2p2nqsd0c76g0\app\resources
+  LD_LIBRARY_PATH              = (unset)
+  LIBRARY_PATH                 = (unset)
+  CUDA_HOME                    = (unset)
+  CUDA_PATH                    = C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4
+  CUDA_VISIBLE_DEVICES         = 1
+  NVIDIA_VISIBLE_DEVICES       = (unset)
+  NVIDIA_DRIVER_CAPABILITIES   = (unset)
+  GGML_CUDA_DEBUG              = 1
+  GGML_VK_VISIBLE_DEVICES      = 1
+  GGML_VK_PIPELINE_CACHE_DEBUG = 1
+  CRISPASR_USE_CUDA_COMPAT     = (unset)
+  CRISPASR_BACKEND             = (unset)
+  CRISPASR_CACHE_DIR           = (unset)
+
+=== ggml backends + devices ===
+ggml_vulkan: Found 1 Vulkan devices:
+ggml_vulkan: 0 = AMD Radeon(TM) 890M Graphics (AMD proprietary driver) | uma: 1 | fp16: 1 | bf16: 1 | fp4: 0 | warp size: 64 | shared memory: 32768 | int dot: 1 | matrix cores: KHR_coopmat
+  registered backends: 2
+    [0] Vulkan (devices: 1)
+    [1] CPU (devices: 1)
+  registered devices : 2
+    [0] igpu   name=Vulkan0 desc=AMD Radeon(TM) 890M Graphics mem=46080/48505 MiB id=?
+    [1] cpu    name=CPU desc=AMD Ryzen AI 9 HX 370 w/ Radeon 890M            mem=30022/64633 MiB id=?
+
+cpu-isa: OK — built for AVX2/FMA/F16C/BMI2/AVX; host CPU supports the build's ISA baseline
+crispasr_init_gpu_backend: using preferred GPU backend: Vulkan0
+qwen3_tts: code_pred CPU-pinned (72 tensors, 432 MB, copy=f32)
+qwen3_tts: kv cache 448 MiB (head_dim=128 max_ctx=4096 n_kv=8 n_layers=28)
+qwen3_tts: loaded C:\Users\zhout\w2\models\crisp_qwen3tts_0.6b\qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf  (talker 28L/1024  code_pred 5L  vocab 3072)
+qwen3_tts: codec: GPU default - loading weights onto Vulkan0
+qwen3_tts: codec loaded from 'C:\Users\zhout\w2\models\crisp_qwen3tts_0.6b\qwen3-tts-tokenizer-12hz.gguf'  (8L d=512/1024  rvq=16)
+qwen3_tts: codec encoder loaded (SEANet+xfmr+RVQ)
+crispasr[qwen3-tts]: codec loaded from 'C:\Users\zhout\w2\models\crisp_qwen3tts_0.6b\qwen3-tts-tokenizer-12hz.gguf'
+crispasr-server: warmup completed in 0 ms
+crispasr-server: backend 'qwen3-tts-customvoice' loaded, model 'C:\Users\zhout\w2\models\crisp_qwen3tts_0.6b\qwen3-tts-12hz-0.6b-customvoice-q8_0.gguf'
+fireredpunc: 12L, d=768, ffn=3072, heads=12, vocab=21128, labels=5
+crispasr_init_gpu_backend: using preferred GPU backend: Vulkan0
+crispasr-server: loaded punctuation model 'C:\Users\zhout\.cache\crispasr\fireredpunc-q4_k.gguf'
+
+crispasr-server: listening on 127.0.0.1:64961
+  POST /inference                  — upload audio (native JSON)
+  POST /v1/audio/transcriptions    — OpenAI-compatible API
+  POST /v1/audio/speech            — TTS (OpenAI-compatible)
+  POST /load                       — hot-swap model
+  GET  /health                     — server status
+  GET  /backends                   — list backends
+  GET  /v1/models                  — model info
+  GET  /v1/voices                  — list voices in --voice-dir
+       /voices, /v1/audio/voices   — aliases (llama-swap compat)
+  POST /v1/voices                  — upload voice file (multipart)
+  DELETE /v1/voices/:name          — delete voice file
+crispasr-server: warning: --voice-dir not set; /v1/voices will return empty and /v1/audio/speech will reject requests with a 'voice' field
+
+crispasr: warning: 'qwen3-tts-customvoice' does not record whether its preset voice belongs to a real person, so no spoken AI
+  disclosure was added. If the voice is that of an identifiable individual, the output is a
+  deep fake under EU AI Act Art. 3(60) and you carry the Art. 50(4) duty to disclose it.
+  Pass --speaker-identity real_person to have CrispASR prepend the disclosure, or
+  --speaker-identity synthetic to silence this.
+qwen3_tts: selected CustomVoice speaker 'ryan' (spk_id=3061)
+crispasr[qwen3-tts]: CustomVoice speaker = 'ryan' (available: aiden, dylan, eric, ono_anna, ryan, serena, sohee, uncle_fu, vivian)
+qwen3_tts[customvoice]: prefill role=3 + bridge=5 + text=7 + final=1 = T=16 trailing=1
+qwen3_tts: produced 32 frames × 16 codebooks = 512 codes
+qwen3_tts: perf — codes=3121 ms  codec=219 ms  total=3340 ms  audio=2.6 s  rtf=1.305
+crispasr-server: synthesized 2.6s audio in 3.35s (RTF=1.31) voice='<startup>' speed=1.00 format=wav model='<unset>' chunks=1 sr=24000Hz
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_device_20260923.driver.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_device_20260923.driver.log
new file mode 100644
index 000000000..344b3fc91
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_device_20260923.driver.log
@@ -0,0 +1,100 @@
+INFO 09-23 20:50:44 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
+INFO 09-23 20:50:44 [patch.py:510] inductor factorable-divisibility patch: installed.
+INFO 09-23 20:50:48 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
+WARNING 09-23 20:50:48 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 13] Permission denied: 'C:\\Users\\zhout\\w2'
+ERROR 09-23 20:50:50 [stage_runtime.py:265] [StageRuntime] Stage initialization failed; shutting down 0 initialized client(s)
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265] Traceback (most recent call last):
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 252, in initialize
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     initialized_clients = self._initialize_stage_replicas(stage_plans, self._stage_init_timeout)
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 589, in _initialize_stage_replicas
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     raise primary_exc
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 550, in _init_group
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     client = self._initialize_replica(
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]              ^^^^^^^^^^^^^^^^^^^^^^^^^
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 817, in _initialize_replica
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     return create_graph_client(
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]            ^^^^^^^^^^^^^^^^^^^^
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\__init__.py", line 19, in create_graph_client
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     return CrispTTSStageClient(metadata, backend, ledger, reservation)
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\crisp_tts.py", line 139, in __init__
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     raise ValueError("this shared-RAM hybrid route is qualified only for the HX370 Radeon 890M")
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265] ValueError: this shared-RAM hybrid route is qualified only for the HX370 Radeon 890M
+{
+  "scope": "complete resident Qwen3-TTS CustomVoice text-to-WAV through Omni StageRuntime/StagePool",
+  "placement": "Radeon Vulkan talker/codec + CPU FP32 code predictor",
+  "warmup_count": 0,
+  "measured_count": 1,
+  "memory_budget": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "demands": {
+      "host_ram": 8589934592
+    }
+  },
+  "status": "failed",
+  "error": "ValueError: this shared-RAM hybrid route is qualified only for the HX370 Radeon 890M",
+  "ledger_after_shutdown": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "reserved": {
+      "host_ram": 0
+    },
+    "owners": [],
+    "quarantined": []
+  }
+}
+Traceback (most recent call last):
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\benchmarks\edge_harness\experiments\probe_omni_crisp_tts.py", line 180, in <module>
+    asyncio.run(main())
+  File "C:\Users\zhout\AppData\Local\Programs\Python\Python312\Lib\asyncio\runners.py", line 195, in run
+    return runner.run(main)
+           ^^^^^^^^^^^^^^^^
+  File "C:\Users\zhout\AppData\Local\Programs\Python\Python312\Lib\asyncio\runners.py", line 118, in run
+    return self._loop.run_until_complete(task)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "C:\Users\zhout\AppData\Local\Programs\Python\Python312\Lib\asyncio\base_events.py", line 691, in run_until_complete
+    return future.result()
+           ^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\benchmarks\edge_harness\experiments\probe_omni_crisp_tts.py", line 100, in main
+    runtime.initialize()
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 278, in initialize
+    raise exc
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 252, in initialize
+    initialized_clients = self._initialize_stage_replicas(stage_plans, self._stage_init_timeout)
+                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 589, in _initialize_stage_replicas
+    raise primary_exc
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 550, in _init_group
+    client = self._initialize_replica(
+             ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 817, in _initialize_replica
+    return create_graph_client(
+           ^^^^^^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\__init__.py", line 19, in create_graph_client
+    return CrispTTSStageClient(metadata, backend, ledger, reservation)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\crisp_tts.py", line 139, in __init__
+    raise ValueError("this shared-RAM hybrid route is qualified only for the HX370 Radeon 890M")
+ValueError: this shared-RAM hybrid route is qualified only for the HX370 Radeon 890M
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_device_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_device_20260923.json
new file mode 100644
index 000000000..d6608dadc
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_device_20260923.json
@@ -0,0 +1,26 @@
+{
+  "scope": "complete resident Qwen3-TTS CustomVoice text-to-WAV through Omni StageRuntime/StagePool",
+  "placement": "Radeon Vulkan talker/codec + CPU FP32 code predictor",
+  "warmup_count": 0,
+  "measured_count": 1,
+  "memory_budget": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "demands": {
+      "host_ram": 8589934592
+    }
+  },
+  "status": "failed",
+  "error": "ValueError: this shared-RAM hybrid route is qualified only for the HX370 Radeon 890M",
+  "ledger_after_shutdown": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "reserved": {
+      "host_ram": 0
+    },
+    "owners": [],
+    "quarantined": []
+  }
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_memory_20260923.driver.log b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_memory_20260923.driver.log
new file mode 100644
index 000000000..2113b9c7d
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_memory_20260923.driver.log
@@ -0,0 +1,100 @@
+INFO 09-23 20:50:44 [patch.py:259] NVFP4 W4A4 weight_scale NaN-clamp: installed.
+INFO 09-23 20:50:44 [patch.py:510] inductor factorable-divisibility patch: installed.
+INFO 09-23 20:50:48 [stage_init_utils.py:768] [stage_init] Set VLLM_WORKER_MULTIPROC_METHOD=spawn
+WARNING 09-23 20:50:48 [stage_init_utils.py:1887] [stage_init] Failed to load transfer config: [Errno 13] Permission denied: 'C:\\Users\\zhout\\w2'
+ERROR 09-23 20:50:50 [stage_runtime.py:265] [StageRuntime] Stage initialization failed; shutting down 0 initialized client(s)
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265] Traceback (most recent call last):
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 252, in initialize
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     initialized_clients = self._initialize_stage_replicas(stage_plans, self._stage_init_timeout)
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 589, in _initialize_stage_replicas
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     raise primary_exc
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 550, in _init_group
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     client = self._initialize_replica(
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]              ^^^^^^^^^^^^^^^^^^^^^^^^^
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 817, in _initialize_replica
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     return create_graph_client(
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]            ^^^^^^^^^^^^^^^^^^^^
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\__init__.py", line 19, in create_graph_client
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     return CrispTTSStageClient(metadata, backend, ledger, reservation)
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]   File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\crisp_tts.py", line 132, in __init__
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265]     raise ResourceUnavailable("TTS weights plus declared state/workspace/headroom exceed reservation")
+
+ERROR 09-23 20:50:50 [stage_runtime.py:265] vllm_omni.engine.resource_ledger.ResourceUnavailable: TTS weights plus declared state/workspace/headroom exceed reservation
+{
+  "scope": "complete resident Qwen3-TTS CustomVoice text-to-WAV through Omni StageRuntime/StagePool",
+  "placement": "Radeon Vulkan talker/codec + CPU FP32 code predictor",
+  "warmup_count": 0,
+  "measured_count": 1,
+  "memory_budget": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "demands": {
+      "host_ram": 4294967296
+    }
+  },
+  "status": "failed",
+  "error": "ResourceUnavailable: TTS weights plus declared state/workspace/headroom exceed reservation",
+  "ledger_after_shutdown": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "reserved": {
+      "host_ram": 0
+    },
+    "owners": [],
+    "quarantined": []
+  }
+}
+Traceback (most recent call last):
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\benchmarks\edge_harness\experiments\probe_omni_crisp_tts.py", line 180, in <module>
+    asyncio.run(main())
+  File "C:\Users\zhout\AppData\Local\Programs\Python\Python312\Lib\asyncio\runners.py", line 195, in run
+    return runner.run(main)
+           ^^^^^^^^^^^^^^^^
+  File "C:\Users\zhout\AppData\Local\Programs\Python\Python312\Lib\asyncio\runners.py", line 118, in run
+    return self._loop.run_until_complete(task)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "C:\Users\zhout\AppData\Local\Programs\Python\Python312\Lib\asyncio\base_events.py", line 691, in run_until_complete
+    return future.result()
+           ^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\benchmarks\edge_harness\experiments\probe_omni_crisp_tts.py", line 100, in main
+    runtime.initialize()
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 278, in initialize
+    raise exc
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 252, in initialize
+    initialized_clients = self._initialize_stage_replicas(stage_plans, self._stage_init_timeout)
+                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 589, in _initialize_stage_replicas
+    raise primary_exc
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 550, in _init_group
+    client = self._initialize_replica(
+             ^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\stage_runtime.py", line 817, in _initialize_replica
+    return create_graph_client(
+           ^^^^^^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\__init__.py", line 19, in create_graph_client
+    return CrispTTSStageClient(metadata, backend, ledger, reservation)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "\\wsl.localhost\Ubuntu\home\zhout\project\edge_infer\vllm-omni-edge\vllm_omni\engine\backends\crisp_tts.py", line 132, in __init__
+    raise ResourceUnavailable("TTS weights plus declared state/workspace/headroom exceed reservation")
+vllm_omni.engine.resource_ledger.ResourceUnavailable: TTS weights plus declared state/workspace/headroom exceed reservation
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_memory_20260923.json b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_memory_20260923.json
new file mode 100644
index 000000000..b31d70312
--- /dev/null
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/omni_crisp_tts_pinned_reject_memory_20260923.json
@@ -0,0 +1,26 @@
+{
+  "scope": "complete resident Qwen3-TTS CustomVoice text-to-WAV through Omni StageRuntime/StagePool",
+  "placement": "Radeon Vulkan talker/codec + CPU FP32 code predictor",
+  "warmup_count": 0,
+  "measured_count": 1,
+  "memory_budget": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "demands": {
+      "host_ram": 4294967296
+    }
+  },
+  "status": "failed",
+  "error": "ResourceUnavailable: TTS weights plus declared state/workspace/headroom exceed reservation",
+  "ledger_after_shutdown": {
+    "capacities": {
+      "host_ram": 17179869184
+    },
+    "reserved": {
+      "host_ram": 0
+    },
+    "owners": [],
+    "quarantined": []
+  }
+}
diff --git a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md
index 84a8bcd43..158a90271 100644
--- a/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md
+++ b/benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md
@@ -1,6 +1,6 @@
 # Native Windows CPU + Radeon 890M Qwen3-TTS scoped whole-model path
 
-The 0.6B CustomVoice model completed two named English **text-to-WAV** requests in [CrispASR v0.8.35](https://github.com/CrispStrobe/CrispASR/releases/tag/v0.8.35), outside Omni. This is a **joint CPU+iGPU result**: the talker and codec were routed to Radeon 890M Vulkan, while the iterative code predictor was explicitly pinned to CPU FP32. It is not an all-Radeon pass or a CPU+iGPU+NPU pass. The [DirectML route failed quality](../qwen_tts_radeon890m_directml/README.md). Neither route establishes Omni streaming or release qualification.
+The 0.6B CustomVoice model completed two named English **text-to-WAV** requests in [CrispASR v0.8.35](https://github.com/CrispStrobe/CrispASR/releases/tag/v0.8.35), outside Omni. A subsequent [bounded Omni whole-session stage](../qwen_tts_omni_hybrid/README.md) now executes the same hybrid route with exact PCM parity on both inputs. This is a **joint CPU+iGPU result**: the talker and codec were routed to Radeon 890M Vulkan, while the iterative code predictor was explicitly pinned to CPU FP32. It is not an all-Radeon pass or a CPU+iGPU+NPU pass. The [DirectML route failed quality](../qwen_tts_radeon890m_directml/README.md). Neither route establishes playable Omni streaming or release qualification.
 
 The host was Windows 11 build 26200 on Ryzen AI 9 HX 370 with Radeon 890M driver `32.0.22018.6001`; an RTX 5090 Laptop was also present. `GGML_VK_VISIBLE_DEVICES=1` exposed only the Radeon adapter, enumerated as `Vulkan0`; `-dev 1` selected the original physical adapter. The [first placement log](crisp_tts_radeon_cp_cpu_20260923.log) records `AMD Radeon(TM) 890M Graphics`, `using preferred GPU backend: Vulkan0`, `code_pred CPU-pinned (72 tensors, 432 MB, copy=f32)`, and `codec: GPU default - loading weights onto Vulkan0`. The [runtime's routing source](https://github.com/CrispStrobe/CrispASR/blob/v0.8.35/src/qwen3_tts.cpp) routes the talker off Vulkan unless `CRISPASR_QWEN3_TTS_VULKAN_NATIVE=1`; that override was set for the passing runs, and the CPU-talkback message was absent. This is placement evidence for the named runtime configuration, not a per-operator profiler trace.
 
diff --git a/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md b/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md
index 3d8b5082a..81c13067c 100644
--- a/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md
+++ b/benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md
@@ -1,6 +1,6 @@
 # E2E check and profiling
 
-All 60 pairings have a disposition. The timing tables below remain the 2026-09-22 measurements. The configuration matrix and affected per-cell reasons incorporate the [2026-09-23 scoped recovery runs](../../../e2e_recovery_20260923/README.md) and [WSL CPU/native Windows expansion](../../../e2e_expansion_20260923/README.md). The expansion adds separate 20-request serial MiniCPM-o text-to-speech profiles on WSL CPU and native Windows GPU, a Spark output-head component profile on the HX370 AMD NPU, complete Spark text-request profiles through standalone llama.cpp and a bounded Omni whole-session stage on native Windows CPU and Radeon 890M, a standalone Qwen3-TTS text-to-WAV profile on native Windows CPU, and a separate hybrid CPU+Radeon 890M Qwen3-TTS profile. Profiling completion does not imply full release qualification.
+All 60 pairings have a disposition. The timing tables below remain the 2026-09-22 measurements. The configuration matrix and affected per-cell reasons incorporate the [2026-09-23 scoped recovery runs](../../../e2e_recovery_20260923/README.md) and [WSL CPU/native Windows expansion](../../../e2e_expansion_20260923/README.md). The expansion adds separate 20-request serial MiniCPM-o text-to-speech profiles on WSL CPU and native Windows GPU, a Spark output-head component profile on the HX370 AMD NPU, complete Spark text-request profiles through standalone llama.cpp and a bounded Omni whole-session stage on native Windows CPU and Radeon 890M, a standalone Qwen3-TTS text-to-WAV profile on native Windows CPU, and a hybrid CPU+Radeon 890M Qwen3-TTS profile now repeated through a bounded Omni whole-session stage. Profiling completion does not imply full release qualification.
 
 ## Configuration × model
 
@@ -12,7 +12,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | PC HX370: CPU / native Windows | scoped E2E: Omni Spark text | scoped E2E: standalone text → WAV | NOT E2E | NOT E2E | synthetic policy pass |
 | PC HX370 + RTX 5090 Laptop / WSL | profile completed | profile completed | scoped E2E: text + image | scoped E2E: text/image/audio → speech | synthetic policy pass |
 | PC HX370 + RTX 5090 Laptop / Windows | profile completed | profile completed | NOT E2E | scoped E2E: text/image/audio → speech | synthetic policy pass |
-| PC HX370 + Radeon 890M / Windows worker | scoped E2E: Omni Spark text | scoped E2E: Radeon talker/codec + CPU predictor | NOT E2E | NOT E2E | NOT E2E |
+| PC HX370 + Radeon 890M / Windows worker | scoped E2E: Omni Spark text | scoped E2E: Omni Radeon talker/codec + CPU predictor | NOT E2E | NOT E2E | NOT E2E |
 | PC HX370 + AMD NPU / Windows worker | NOT E2E (output-head component passed) | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
 | PC CPU+iGPU+NPU +/- discrete GPU, joint model execution | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
 | PC Snapdragon X Elite CRD / AI Hub | NOT E2E | NOT E2E | NOT E2E | NOT E2E | NOT E2E |
@@ -46,7 +46,7 @@ NOT E2E means no complete device/model workload was verified. “Scoped E2E” i
 | PC HX370 + RTX 5090 Laptop / Windows | MiniCPM-o 4.5 | SCOPED_E2E | The verified checkpoint completed native Windows three-stage BF16 text, one synthetic image, and one synthetic audio input to text+speech in separate constrained plans with 5 GiB thinker host offload and explicit 2,048/512-token limits. Text and all 32 thinker token IDs matched WSL CUDA; a 5.2 s WAV was nonzero. The red-square image produced the same description and 27 thinker token IDs as WSL plus a nonzero 5.2 s WAV in 17.15 s. The 2 s/440 Hz tone produced the same whistling description and nine IDs as WSL plus a nonzero 2.16 s WAV in 8.26 s. The initial text run needed 900/600 s startup limits after a 300 s timeout; a separate 1-warmup/20-measured serial text run had request wall p50 18.38 s/p95 18.73 s, with sampled host available RAM as low as 0.78 GB and pagefile use up to 11.17 GB. These synthetic inputs do not establish general image/audio or speech quality. | Check real image/audio suites, video input, speech intelligibility/alignment, concurrency, interruption, loading-peak admission and sustained thermal/power behavior before broader qualification; [native evidence](../../../e2e_expansion_20260923/README.md). |
 | PC HX370 + RTX 5090 Laptop / Windows | InternVLA-A1 | SYNTHETIC_POLICY_PASS | The Place_Markpen checkpoint strict-loaded from a UNC path and produced finite `[1, 50, 32]` actions on synthetic zero observations/noise. Policy/output CUDA placement was checked; one native forward took 1.12 s, with 6.68 GB peak allocated. Relative L2 difference from WSL CUDA was 0.83% (no task tolerance). | Validate real observations/reference actions, units/timestamps, memory and repeated latency before task qualification; [native evidence](../../../e2e_expansion_20260923/README.md). |
 | PC HX370 + Radeon 890M / Windows worker | Spark-X2.5 | SCOPED_E2E_OMNI_COMPLETE_REQUEST | The same pinned Spark-X2.5-1.7B Q4_K_M GGUF completed text prefill/decode through a bounded Omni whole-session stage with llama.cpp Vulkan1. The log identifies Radeon 890M and 29/29 layers offloaded there, rather than RTX 5090; an intentionally wrong device name was refused. `Paris` and `1511` passed; one warmup plus 20 serial complete requests were 20/20 correct, nearest-rank wall p50/p95 2.10/2.14 s. A public AsyncOmni request passed; an oversized context was rejected with a later correct request; in-flight cancellation produced no stale output and released the ledger. The prior standalone run was 2.18/2.20 s. No streaming, restart-after-cancel, GPU-memory peak or sustained power/thermal check is claimed. | Qualify public configuration, broader text/token quality, streaming and cancellation recovery, shared-RAM/GPU peak and sustained use; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md), [standalone evidence](../../../e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md). |
-| PC HX370 + Radeon 890M / Windows worker | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_STANDALONE_HYBRID | The Q8_0 CustomVoice GGUF and F16 codec GGUF completed two named text-to-WAV requests in standalone CrispASR on native Windows with Vulkan0 placed on Radeon 890M for the talker/codec and the 432 MB code predictor explicitly pinned to CPU FP32. Whisper tiny.en transcribed both sentences exactly (WER 0 as an intelligibility proxy). One warmup plus 20 serial resident-server requests for one repeated prompt had nearest-rank client wall p50/p95 3.310/3.345 s for identical 2.56 s PCM (RTF 1.29/1.31). The default Vulkan configuration actually fell back to CPU; forcing all stages to Vulkan ran away to the 360-frame cap. A separate DirectML route failed waveform quality. No Omni binding, playable streaming, loading peak, broad quality or sustained power/thermal run was verified. | Bind the whole hybrid session with explicit CPU+iGPU placement/admission and cancellation; validate quality, real-time playback and full-chain benefit before broader qualification; [joint backend evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md), [failed DirectML route](../../../e2e_expansion_20260923/evidence/qwen_tts_radeon890m_directml/README.md). |
+| PC HX370 + Radeon 890M / Windows worker | Qwen3-TTS 0.6B CustomVoice | SCOPED_E2E_OMNI_COMPLETE_REQUEST_HYBRID | The pinned Q8_0 CustomVoice talker, F16 codec and punctuation GGUFs completed two named text-to-WAV requests through an Omni whole-session CrispASR stage on native Windows. The runtime log identifies Radeon 890M Vulkan0 for talker/codec and the 432 MB code predictor pinned to CPU FP32; this is not all-Radeon or NPU execution. Both PCM hashes exactly match the earlier standalone outputs. Whisper tiny.en transcribed the first Omni WAV and both standalone WAVs exactly (WER 0 as an intelligibility proxy). One warmup plus 20 serial complete Omni requests for 2.56 s of audio had identical PCM and nearest-rank wall p50/p95 3.298/3.336 s (RTF 1.29/1.30). A public AsyncOmni audio request passed; cancellation drained the worker without stale output or remaining reservation. The earlier standalone profile was 3.310/3.345 s. The default Vulkan configuration previously fell back to CPU, all-Vulkan ran away to a 360-frame cap, and DirectML failed waveform quality. No playable stream, post-cancel restart, loading peak, broad speech quality or sustained power/thermal run is claimed. | Validate playable streaming, restart/recovery, broader voice quality, shared-RAM/GPU peak, concurrency and sustained real-time behavior before wider qualification; [Omni raw evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md), [standalone evidence](../../../e2e_expansion_20260923/evidence/qwen_tts_radeon890m_joint/README.md), [failed DirectML route](../../../e2e_expansion_20260923/evidence/qwen_tts_radeon890m_directml/README.md). |
 | PC HX370 + Radeon 890M / Windows worker | Qwen3.8-27B | NOT_E2E_PROFILEABLE | FP32 448x448 vision tower: DirectML 85/113 nodes; normalized max error 9.095e-06 versus torch CPU. Historical component path, not full 27B or current graph-stage image-to-text integration. | Recover the hashed tower bundle and qualify its current Omni graph handoff with the language stage. |
 | PC HX370 + Radeon 890M / Windows worker | MiniCPM-o 4.5 | NOT_E2E_PROFILEABLE | No placement- and quality-verified whole-model graph bundle/stateful adapter for this model on this AMD route. Small backend graphs do not establish model support. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
 | PC HX370 + Radeon 890M / Windows worker | InternVLA-A1 | NOT_E2E_PROFILEABLE | No placement- and quality-verified whole-model graph bundle/stateful adapter for this model on this AMD route. Small backend graphs do not establish model support. | Supply and validate the missing matching artifact/backend, then run the complete specified workload. |
diff --git a/docs/design/omni_edge_backends.md b/docs/design/omni_edge_backends.md
index cb08163d6..aa611aef3 100644
--- a/docs/design/omni_edge_backends.md
+++ b/docs/design/omni_edge_backends.md
@@ -3,12 +3,17 @@
 The first backend is `external.graph.v1`: one complete, stateless graph operation
 per request, with a fixed input bucket and copied host tensors. A second backend,
 `external.llamacpp.text.v1`, owns one pinned GGUF whole-model server process and
-returns complete text requests. Both use the normal `Omni` / `AsyncOmni` APIs,
+returns complete text requests. `external.crisp.tts.v1` owns a pinned
+CustomVoice GGUF talker, codec and local hybrid CPU+Radeon process, returning
+complete PCM audio. All use the normal `Omni` / `AsyncOmni` APIs,
 `PipelineConfig`, `StageRuntime`, `StagePool` and orchestrator. Native vLLM stages
 retain their model execution, KV cache, sampling and batching. No accelerator is
 mandatory. The llama.cpp backend currently uses the complete-request graph
 control path for admission and acknowledgement; it is not an ONNX graph and does
 not yet support incremental token events or restart after active cancellation.
+The CrispASR route likewise lacks playable incremental PCM and post-cancel
+restart; its current placement gate is specific to the measured HX370 Radeon
+890M shared-RAM configuration.
 
 ## Boundaries
 
@@ -57,6 +62,8 @@ memory-accounting and unavailable-device gaps remain documented repair items.
 
 This section preserves the 2026-09-22 audit. For the latest per-cell disposition and 2026-09-23 follow-up evidence, use the [current 60-cell E2E matrix](../../benchmarks/edge_harness/results/e2e_profiling_20260922/evidence/summary/README.md). The HX370 AMD NPU now has a placement- and quality-verified **Spark 1.7B output-head component** through Omni, while complete Spark generation on that route remains unverified; see the [component record](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_amd_npu_output_head/README.md). Native Windows CPU and Radeon 890M completed [scoped whole-model Spark text requests](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_radeon890m_llamacpp/README.md) with the same Q4_K_M GGUF in standalone llama.cpp, then [complete requests through a bounded Omni stage](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md). The Omni route verifies pinned artifacts, actual placement, one-slot memory admission, terminal events and cancellation drain. Incremental streaming, restart after cancellation, memory peaks and broader quality remain open. These results advance the M3 hardware/backend evidence without claiming a joint multi-accelerator plan.
 
+The [Qwen3-TTS CPU+Radeon whole-session path](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md) also completed two named text-to-audio inputs and 20 serial Omni requests with exact PCM parity to its standalone record. It verifies pinned executable and weights, logged hybrid placement, one-slot shared-RAM admission, terminal audio and cancellation drain. Playable streaming, post-cancel restart, loading/iGPU peaks, sustained use and broader speech quality remain open; RTF still exceeds 1. No joint CPU+iGPU+NPU claim follows from this result.
+
 **All 60 named pairings were checked on 2026-09-22** through actual runs,
 raw-record review or artifact/runtime preflight. This is not 60 successful
 executions. Each cell links to its evidence, finding and next required step.
diff --git a/docs/design/omni_edge_e2e_roadmap.md b/docs/design/omni_edge_e2e_roadmap.md
index 46a137e26..fd44ca377 100644
--- a/docs/design/omni_edge_e2e_roadmap.md
+++ b/docs/design/omni_edge_e2e_roadmap.md
@@ -133,6 +133,8 @@ outputs need reference analysis; timing alone cannot identify their cause.
 
 The 2026-09-23 [Spark GGUF Omni follow-up](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/spark_omni_llamacpp/README.md) adds a separate native Windows CPU/Radeon 890M whole-session option. Its pinned 1.7B Q4_K_M artifact passed two named complete text requests and 20 serial inventory requests per device through `StageRuntime`/`StagePool`; public `AsyncOmni` short requests passed on both devices. It verifies CPU or Vulkan1 placement, explicit 4 GiB host reservation, terminal event/acknowledgement and cancellation drain. This does not change the vLLM M0 acceptance or establish M1 mobile generation. Its next gates are incremental output, post-cancel restart, broader token/quality parity, loading/device-memory peaks, concurrency and sustained power/thermal behavior.
 
+The [Qwen3-TTS Omni hybrid follow-up](../../benchmarks/edge_harness/results/e2e_expansion_20260923/evidence/qwen_tts_omni_hybrid/README.md) adds one complete-request stage for the native Windows HX370 CPU+Radeon 890M route. A pinned Q8_0 talker and F16 codec produced two named PCM outputs exactly matching the prior standalone backend; one warmup plus 20 serial requests measured 3.298/3.336 s p50/p95 for 2.56 s of audio. Public `AsyncOmni` audio, cancellation drain and explicit 8 GiB host admission passed. This is progress toward M2 integration, not M2 acceptance: RTF remains above 1, and playable chunk streaming, post-cancel restart, loading/iGPU peaks, broader voice quality and 30-minute thermal behavior remain open. No NPU participated.
+
 VLA qualification here ends at action data. Physical robot control, collision
 avoidance and actuator safety belong to the external controller.
 
diff --git a/vllm_omni/config/pipeline_registry.py b/vllm_omni/config/pipeline_registry.py
index 1f92d4a7d..b27cc8f50 100644
--- a/vllm_omni/config/pipeline_registry.py
+++ b/vllm_omni/config/pipeline_registry.py
@@ -111,7 +111,11 @@ from vllm_omni.model_executor.models.qwen3_omni.pipeline import (
     QWEN3_OMNI_THINKER_ONLY_PIPELINE,
     resolve_qwen3_omni_pipeline,
 )
-from vllm_omni.model_executor.models.qwen3_tts.pipeline import QWEN3_TTS_PIPELINE, QWEN3_TTS_TALKER_ONLY_PIPELINE
+from vllm_omni.model_executor.models.qwen3_tts.pipeline import (
+    QWEN3_TTS_CRISP_HYBRID_PIPELINE,
+    QWEN3_TTS_PIPELINE,
+    QWEN3_TTS_TALKER_ONLY_PIPELINE,
+)
 from vllm_omni.model_executor.models.spark2_5.pipeline import (
     SPARK2_5_GGUF_TEXT_PIPELINE,
     SPARK2_5_PIPELINE,
@@ -144,6 +148,7 @@ OMNI_PIPELINES: dict[str, PipelineConfig | PipelineResolverFunc] = {
     "qwen3_omni_moe_thinker_only": QWEN3_OMNI_THINKER_ONLY_PIPELINE,
     "qwen3_tts": QWEN3_TTS_PIPELINE,
     "qwen3_tts_talker_only": QWEN3_TTS_TALKER_ONLY_PIPELINE,
+    "qwen3_tts_crisp_hybrid": QWEN3_TTS_CRISP_HYBRID_PIPELINE,
     # Single-stage text: an AR decoder with no companion stage. It lives in
     # the same registry as the multi-stage pipelines so that one code path
     # resolves every model (see the M0 local text mode).
diff --git a/vllm_omni/engine/backends/__init__.py b/vllm_omni/engine/backends/__init__.py
index fb69290b9..34cadbf96 100644
--- a/vllm_omni/engine/backends/__init__.py
+++ b/vllm_omni/engine/backends/__init__.py
@@ -13,4 +13,8 @@ def create_graph_client(metadata, config, ledger, reservation):
         from .llamacpp import LlamaCppTextStageClient
 
         return LlamaCppTextStageClient(metadata, backend, ledger, reservation)
+    if name == "external.crisp.tts.v1":
+        from .crisp_tts import CrispTTSStageClient
+
+        return CrispTTSStageClient(metadata, backend, ledger, reservation)
     raise ValueError(f"unsupported complete-request stage backend: {name!r}")
diff --git a/vllm_omni/engine/backends/crisp_tts.py b/vllm_omni/engine/backends/crisp_tts.py
new file mode 100644
index 000000000..e0eeab3e1
--- /dev/null
+++ b/vllm_omni/engine/backends/crisp_tts.py
@@ -0,0 +1,352 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Bounded complete-request Qwen3-TTS stage backed by local CrispASR.
+
+CrispASR owns the talker, code predictor and codec state. Omni owns the
+reservation, process, request identity, terminal audio and output acknowledgement.
+This backend does not advertise playable incremental audio or post-cancel restart.
+"""
+
+from __future__ import annotations
+
+import asyncio
+import dataclasses
+import hashlib
+import io
+import json
+import os
+import re
+import socket
+import subprocess
+import time
+import urllib.error
+import urllib.request
+import uuid
+import wave
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from omni_stage_contracts import StageEvent, StageRequest
+from vllm.outputs import CompletionOutput
+
+from vllm_omni.engine.resource_ledger import ResourceUnavailable
+from vllm_omni.engine.stage_client import StageClientBase
+from vllm_omni.outputs import OmniRequestOutput
+
+
+def _sha256(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as source:
+        for chunk in iter(lambda: source.read(8 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _local_port() -> int:
+    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as sock:
+        sock.bind(("127.0.0.1", 0))
+        return int(sock.getsockname()[1])
+
+
+def _post_wav(url: str, text: str, seed: int, *, timeout: float, max_bytes: int) -> bytes:
+    payload = json.dumps({"input": text, "response_format": "wav", "seed": seed}).encode()
+    request = urllib.request.Request(
+        url, data=payload, headers={"Content-Type": "application/json"}, method="POST"
+    )
+    with urllib.request.urlopen(request, timeout=timeout) as response:
+        if response.status != 200:
+            raise RuntimeError(f"CrispASR returned HTTP {response.status}")
+        wav_bytes = response.read(max_bytes + 1)
+    if len(wav_bytes) > max_bytes:
+        raise ResourceUnavailable("CrispASR WAV exceeds admitted output bound")
+    return wav_bytes
+
+
+def _decode_wav(wav_bytes: bytes, max_audio_s: float) -> tuple[torch.Tensor, bytes, int]:
+    with wave.open(io.BytesIO(wav_bytes), "rb") as wav:
+        channels, width, sample_rate, frames = (
+            wav.getnchannels(), wav.getsampwidth(), wav.getframerate(), wav.getnframes()
+        )
+        if (channels, width, sample_rate) != (1, 2, 24000) or not 0 < frames <= max_audio_s * sample_rate:
+            raise ValueError("CrispASR WAV violates 24 kHz mono PCM16 or duration contract")
+        pcm = wav.readframes(frames)
+    if len(pcm) != frames * 2:
+        raise ValueError("CrispASR returned truncated PCM")
+    audio = torch.from_numpy(np.frombuffer(pcm, dtype="<i2").astype(np.float32) / 32768.0)
+    if not torch.isfinite(audio).all() or float(torch.sqrt(torch.mean(audio.square()))) <= 1e-5:
+        raise ValueError("CrispASR returned silent or invalid audio")
+    return audio, pcm, frames
+
+
+class CrispTTSStageClient(StageClientBase):
+    def __init__(self, metadata, config: dict, ledger, reservation) -> None:
+        for name, value in vars(metadata).items():
+            setattr(self, name, value)
+        self.stage_type = "graph"  # Omni's complete-request control path.
+        self._ledger, self._reservation = ledger, reservation
+        self._generation = uuid.uuid4().hex
+        self._proc: subprocess.Popen | None = None
+        self._log_stream = None
+        self._closed = False
+        self._active: str | None = None
+        self._task: asyncio.Task | None = None
+        self._output: OmniRequestOutput | None = None
+        self._epoch = 0
+        self._memory_pool = str(config.get("memory_pool", "host_ram"))
+        self._max_text_bytes = int(config.get("max_text_bytes", 4096))
+        self._max_wav_bytes = int(config.get("max_wav_bytes", 4 << 20))
+        self._max_audio_s = float(config.get("max_audio_s", 10))
+        self._request_timeout_s = float(config.get("request_timeout_s", 180))
+        self._voice = str(config.get("voice", "ryan"))
+        self._seed = int(config.get("seed", 42))
+        self._port = int(config.get("port") or _local_port())
+        try:
+            if (
+                self._max_text_bytes <= 0 or self._max_wav_bytes <= 0 or self._max_audio_s <= 0
+                or self._request_timeout_s <= 0 or not 0 < self._port < 65536
+                or self._seed < 1 or not re.fullmatch(r"[a-z_]+", self._voice)
+            ):
+                raise ValueError("invalid CrispASR text, WAV, voice, seed, timeout or port bound")
+            if self._memory_pool not in reservation.demands:
+                raise ResourceUnavailable("CrispASR memory pool absent from stage reservation")
+            demand = reservation.demands[self._memory_pool]
+            if self._max_wav_bytes > demand:
+                raise ResourceUnavailable("CrispASR WAV bound exceeds stage reservation")
+            self._binary = Path(config["server_bin"]).resolve(strict=True)
+            self._talker = Path(config["talker_file"]).resolve(strict=True)
+            self._codec = Path(config["codec_file"]).resolve(strict=True)
+            self._punc = Path(config["punc_file"]).resolve(strict=True)
+            expected = {
+                "server_sha256": (self._binary, str(config["server_sha256"]).lower()),
+                "talker_sha256": (self._talker, str(config["talker_sha256"]).lower()),
+                "codec_sha256": (self._codec, str(config["codec_sha256"]).lower()),
+                "punc_sha256": (self._punc, str(config["punc_sha256"]).lower()),
+            }
+            for label, (path, digest) in expected.items():
+                if _sha256(path) != digest:
+                    raise ValueError(f"CrispASR {label} differs from the declared artifact hash")
+            overhead = int(config["memory_overhead_bytes"])
+            size = self._talker.stat().st_size + self._codec.stat().st_size + self._punc.stat().st_size
+            if overhead <= 0 or size + overhead > demand:
+                raise ResourceUnavailable("TTS weights plus declared state/workspace/headroom exceed reservation")
+            self._expected_gpu_name = str(config["expected_gpu_name"])
+            if not self._expected_gpu_name:
+                raise ValueError("hybrid TTS stage requires expected_gpu_name")
+            if self._expected_gpu_name != "AMD Radeon(TM) 890M Graphics" or str(
+                config.get("ggml_vk_visible_devices", "1")
+            ) != "1":
+                raise ValueError("this shared-RAM hybrid route is qualified only for the HX370 Radeon 890M")
+            self._log_path = Path(config["log_file"]).resolve()
+            self._log_path.parent.mkdir(parents=True, exist_ok=True)
+            self._log_stream = self._log_path.open("wb")
+            env = os.environ.copy()
+            env["GGML_VK_VISIBLE_DEVICES"] = str(config.get("ggml_vk_visible_devices", "1"))
+            env["CRISPASR_QWEN3_TTS_VULKAN_NATIVE"] = "1"
+            env["CRISPASR_QWEN3_TTS_CP_BACKEND"] = "cpu-f32"
+            command = [
+                str(self._binary), "--server", "--host", "127.0.0.1", "--port", str(self._port),
+                "--backend", "qwen3-tts-customvoice", "-m", str(self._talker),
+                "--codec-model", str(self._codec), "--voice", self._voice,
+                "--punc-model", str(self._punc),
+                "--gpu-backend", "vulkan", "-dev", "1", "--no-flash-attn",
+                "-n", "96", "--seed", str(self._seed), "--verbose",
+            ]
+            flags = subprocess.CREATE_NO_WINDOW if os.name == "nt" else 0
+            self._proc = subprocess.Popen(
+                command, stdout=self._log_stream, stderr=subprocess.STDOUT,
+                env=env, creationflags=flags,
+            )
+            base = f"http://127.0.0.1:{self._port}"
+            deadline = time.monotonic() + float(config.get("start_timeout_s", 180))
+            while True:
+                if self._proc.poll() is not None:
+                    raise RuntimeError(f"CrispASR exited during load; see {self._log_path}")
+                try:
+                    with urllib.request.urlopen(base + "/health", timeout=1) as response:
+                        if response.status == 200:
+                            break
+                except (urllib.error.URLError, TimeoutError, OSError):
+                    pass
+                if time.monotonic() >= deadline:
+                    raise TimeoutError(f"CrispASR did not become ready; see {self._log_path}")
+                time.sleep(0.2)
+            log_text = self._log_path.read_text(encoding="utf-8", errors="replace")
+            placement = (
+                f"0 = {self._expected_gpu_name}" in log_text
+                and "using preferred GPU backend: Vulkan0" in log_text
+                and "code_pred CPU-pinned" in log_text
+                and "codec: GPU default - loading weights onto Vulkan0" in log_text
+                and "qwen3_tts: loaded " in log_text
+                and f"loaded punctuation model '{self._punc}'" in log_text
+                and "falling back to CPU" not in log_text
+            )
+            if not placement:
+                raise RuntimeError("REFUSE_DEVICE_PLACEMENT: CrispASR hybrid route differs from declared CPU+Radeon")
+            import psutil
+
+            self._loaded_rss_bytes = int(psutil.Process(self._proc.pid).memory_info().rss)
+            if self._loaded_rss_bytes > demand:
+                raise ResourceUnavailable("loaded CrispASR RSS exceeds stage reservation")
+            self._base_url = base
+            self.execution_plan = {
+                "backend": "external.crisp.tts.v1",
+                "stage_id": self.stage_id,
+                "worker_generation": self._generation,
+                "worker_pid": self._proc.pid,
+                "artifact_sha256": {label: digest for label, (_, digest) in expected.items()},
+                "placement": "Radeon Vulkan0 talker/codec + CPU FP32 code predictor",
+                "expected_gpu_name": self._expected_gpu_name,
+                "ggml_vk_visible_devices": env["GGML_VK_VISIBLE_DEVICES"],
+                "loaded_rss_bytes": self._loaded_rss_bytes,
+                "reserved_bytes": dict(reservation.demands),
+                "memory_overhead_bytes": overhead,
+                "max_text_bytes": self._max_text_bytes,
+                "max_wav_bytes": self._max_wav_bytes,
+                "max_audio_s": self._max_audio_s,
+                "voice": self._voice,
+                "seed": self._seed,
+                "request_capacity": 1,
+                "stateful_session": "CrispASR-owned; complete WAV per request",
+                "evidence": "B",
+            }
+        except BaseException:
+            self.shutdown()
+            raise
+
+    async def add_request_async(self, request_id: str, prompt: Any, params: Any = None) -> None:
+        self.check_health()
+        if self._active is not None:
+            raise ResourceUnavailable("CrispASR stage has an unacknowledged request; capacity is one")
+        if not isinstance(prompt, dict) or not isinstance(prompt.get("text"), str) or not prompt["text"]:
+            raise ValueError("TTS prompt must contain nonempty text")
+        text = prompt["text"]
+        if len(text.encode("utf-8")) > self._max_text_bytes:
+            raise ResourceUnavailable("TTS text exceeds admitted request bound")
+        if prompt.get("voice", self._voice) != self._voice or int(prompt.get("seed", self._seed)) != self._seed:
+            raise ValueError("TTS voice and seed must match the pinned stage configuration")
+        self._epoch += 1
+        request = StageRequest(request_id, self.stage_id, self._epoch, self._generation)
+        self._active = request_id
+        self._task = asyncio.create_task(self._run(request, text), name=f"crisp-tts-{request_id}")
+
+    async def _run(self, request: StageRequest, text: str) -> None:
+        try:
+            started = time.perf_counter()
+            wav = await asyncio.to_thread(
+                _post_wav, self._base_url + "/v1/audio/speech", text, self._seed,
+                timeout=self._request_timeout_s, max_bytes=self._max_wav_bytes,
+            )
+            audio, pcm, frames = _decode_wav(wav, self._max_audio_s)
+            wall_s = time.perf_counter() - started
+            completion = CompletionOutput(0, "", [], None, None, finish_reason="stop")
+            completion.multimodal_output = {"audio": audio, "sr": 24000}
+            event = StageEvent(
+                request.request_id, self.stage_id, request.epoch, 1, "audio",
+                self._generation, terminal=True,
+            )
+            output = OmniRequestOutput(
+                request_id=request.request_id,
+                prompt=text,
+                stage_id=self.stage_id,
+                final_output_type="audio",
+                outputs=[completion],
+                _custom_output={
+                    "stage_event": dataclasses.asdict(event),
+                    "audio_metadata": {
+                        "sample_rate_hz": 24000,
+                        "channels": 1,
+                        "pcm_frames": frames,
+                        "duration_s": frames / 24000,
+                        "pcm_sha256": hashlib.sha256(pcm).hexdigest(),
+                    },
+                },
+                metrics={"crisp_tts_wall_s": wall_s},
+            )
+        except asyncio.CancelledError:
+            raise
+        except Exception as exc:
+            output = OmniRequestOutput.from_error(request.request_id, str(exc))
+            output.stage_id = self.stage_id
+        if not self._closed and self._epoch == request.epoch:
+            loop = asyncio.get_running_loop()
+            output._stage_release = lambda: loop.call_soon_threadsafe(
+                self.acknowledge, request.request_id, request.epoch, request.worker_generation
+            )
+            self._output = output
+
+    def get_graph_output_nowait(self) -> OmniRequestOutput | None:
+        output, self._output = self._output, None
+        return output
+
+    def acknowledge(self, request_id: str, epoch: int, generation: str) -> None:
+        if (
+            self._active == request_id and epoch == self._epoch
+            and generation == self._generation
+            and (self._task is None or self._task.done())
+        ):
+            self._active = None
+            self._task = None
+
+    async def abort_requests_async(self, request_ids: list[str]) -> None:
+        if self._active not in request_ids:
+            return
+        self._epoch += 1
+        self._output = None
+        if self._task is not None and not self._task.done():
+            drained = await asyncio.to_thread(self._terminate)
+            self._closed = True
+            done, _ = await asyncio.wait({self._task}, timeout=5)
+            if done:
+                self._ledger.release(self._reservation, drained=drained)
+            else:
+                self._ledger.release(self._reservation, drained=False)
+                self._task.add_done_callback(
+                    lambda task: self._ledger.release(
+                        self._reservation, drained=drained and not task.cancelled()
+                    )
+                )
+        self._active = None
+
+    def _terminate(self) -> bool:
+        proc = self._proc
+        try:
+            if proc is not None and proc.poll() is None:
+                try:
+                    proc.terminate()
+                except ProcessLookupError:
+                    pass
+                try:
+                    proc.wait(timeout=5)
+                except subprocess.TimeoutExpired:
+                    proc.kill()
+                    proc.wait(timeout=5)
+        except OSError:
+            return False
+        finally:
+            if self._log_stream is not None:
+                self._log_stream.close()
+                self._log_stream = None
+        return proc is None or proc.poll() is not None
+
+    def check_health(self) -> None:
+        if self._closed or self._proc is None or self._proc.poll() is not None:
+            from vllm.v1.engine.exceptions import EngineDeadError
+
+            raise EngineDeadError()
+
+    async def collective_rpc_async(self, method, timeout=None, args=(), kwargs=None):
+        raise NotImplementedError(f"CrispASR backend does not implement collective RPC {method}")
+
+    def shutdown(self) -> None:
+        self._closed = True
+        self._output = None
+        drained = self._terminate()
+        task = self._task
+        if task is not None and not task.done():
+            self._ledger.release(self._reservation, drained=False)
+            task.add_done_callback(
+                lambda done: self._ledger.release(self._reservation, drained=drained and not done.cancelled())
+            )
+        else:
+            self._ledger.release(self._reservation, drained=drained)
diff --git a/vllm_omni/model_executor/models/qwen3_tts/pipeline.py b/vllm_omni/model_executor/models/qwen3_tts/pipeline.py
index 3b25dabb6..6dce12d50 100644
--- a/vllm_omni/model_executor/models/qwen3_tts/pipeline.py
+++ b/vllm_omni/model_executor/models/qwen3_tts/pipeline.py
@@ -83,3 +83,20 @@ QWEN3_TTS_TALKER_ONLY_PIPELINE = PipelineConfig(
         ),
     ),
 )
+
+
+# Whole-session hybrid fallback for the pinned CustomVoice GGUF artifact.
+# CrispASR owns talker, iterative code predictor and codec within one backend;
+# graph is only the Omni complete-request admission and acknowledgement path.
+QWEN3_TTS_CRISP_HYBRID_PIPELINE = PipelineConfig(
+    model_type="qwen3_tts_crisp_hybrid",
+    stages=(
+        StagePipelineConfig(
+            stage_id=0,
+            model_stage="audio",
+            execution_type=StageExecutionType.GRAPH,
+            final_output=True,
+            final_output_type="audio",
+        ),
+    ),
+)