[dengxibo]

Edge online voices in v0.2.9 fail because the adapter still generates the v1 SHA-256 token. This updates the handshake to Sec-MS-GEC v2 using the implementation in edge-tts 7.3.0, while preserving local natural voices and the existing SAPI event conversion.

Changes

• Port the modified SHA-256 and CS64 digest, with integer FILETIME rounding and the existing server Date clock correction.
• Use Sec-MS-GEC-Version=2-143.0.3650.75 and the matching User-Agent.
• Add Origin, MUID cookie, cache-control and language headers only for the official Edge WebSocket endpoint. Do not generate GEC for URLs without the placeholder.
• Add offline reference comparisons and a direct-DLL SAPI smoke test that needs no engine registration.
• Add /p:NaturalVoiceRegisterOutput=false for builds without registration; the existing default remains enabled.

The digest port and pinned reference retain edge-tts's LGPLv3 attribution/license in tests/reference/; the original project code retains its existing MIT license.

Validation

Based on v0.2.9 (7194eca5), which currently matches upstream master:

• Release x64 and x86 builds succeeded, using v145 with ATL and the x64 host tools.
• All 318 C++ token vectors matched edge-tts 7.3.0, including five-minute boundaries and uint64 extremes.
• Edge en-US-AriaNeural worked in both x64 and x86: nonzero PCM, 8 word boundaries, 2 sentence boundaries and 1 bookmark.
• Edge zh-CN-XiaoxiaoNeural worked in x64: nonzero PCM, 6 word boundaries, 2 sentence boundaries, 1 bookmark and 47 visemes.
• Local Aria worked in x64: nonzero PCM, 8 word boundaries, 1 bookmark and 31 visemes.
• Cancellation after 100 ms returned S_OK in approximately 125 ms.
• AzureSpeechSDKShim and TtsApplication x64 builds succeeded.

No installer, registration or SDK patcher was run during validation. Azure subscription endpoints, XP runtime and ARM64 were not tested. See tests/README.md for reproducible commands and test scope.