Map FlowHigh compatibility feasibility

[Tinnci]

Goal

Map whether FlowHigh can become a package-owned compatibility backend without adding heavyweight baseline dependencies or implicit network access.

Source facts to verify

• Repository: https://github.com/resemble-ai/flowhigh
• License: MIT
• Target sample rate: 48 kHz
• Core method: single-step flow matching with transformer and BigVGAN/vocoder components
• Public usage: FlowHighSR.from_pretrained(...).generate(wav, sr_in, target_sr)
• Current checkpoint distribution appears to rely on Google Drive links rather than a hashable provider manifest.

Acceptance criteria

• Identify exact required checkpoint/config files and their licenses.
• Determine whether weights can be mirrored into a managed manifest with stable size/SHA256 metadata.
• Map preprocessing, target-rate behavior, and vocoder boundaries.
• Confirm whether CPU/offline smoke validation is practical.
• Decide whether the next step is external wrapper, self-contained compat backend, or continued watchlist.

Non-goals

• Do not implement the backend in this issue.
• Do not add FlowHigh dependencies to the default package.
• Do not add unverified weights to tests or release artifacts.