Codec roadmap: Python bridge serialization + feature detection

[bbopen]

Goal

Implement Python bridge serialization for the SciPy/Torch/Sklearn envelopes with feature detection, per docs/codec-roadmap.md.

Scope

• Emit SciPy/Torch/Sklearn envelopes matching the documented spec.
• Add feature detection (module availability) surfaced in bridge metadata.
• Provide explicit errors for unsupported formats or missing deps.
• Honor opt-in flags for lossy/CPU conversions (no implicit GPU->CPU or dense conversions).
• Keep JSON fallback opt-in only via TYWRAP_CODEC_FALLBACK=json.

Acceptance criteria

• Python bridge emits the agreed envelopes for all targets.
• Metadata surfaces availability (scipy/torch/sklearn) and matches behavior.
• Errors are explicit when dependencies/encodings are unsupported.
• Behavior matches the documented spec and tests.

References

• docs/codec-roadmap.md

[bbopen]

Roadmap updated in docs/codec-roadmap.md: Arrow defaults + explicit GPU opt-in + payload scaling plan are now locked. This issue remains the Python-bridge implementation track (serialization + feature detection) once #24 lands.

[bbopen]

Resolved by #35 and #37 (bridge serialization + feature detection + Arrow DX).