feat(bindings): who-said-what format parity for Python and C FFI

[ekhodzitsky]

Closes the language-parity gap for who-said-what surfaces: the canonical DiarizationResult v1 is now projectable to every format from Python and C, not just the CLI.

C FFI

• New polyvoice_pipeline_run_format(..., format, ...) — renders the result as JSON / RTTM / SRT / VTT / TXT via a new PolyvoiceFormat (polyvoice_format_t) selector. Same contract as polyvoice_pipeline_run otherwise; RTTM uses the fixed file id audio; unknown formats return InvalidArg before the pipeline is touched.
• Existing ABI untouched (additive only). include/polyvoice.h and the module contract updated (also fixed stale polyvoice_pipeline_free naming → polyvoice_pipeline_destroy, ABI v2 → v3).

Python

• Typed DiarizationResult: from_json() / to_json() / to_rttm(file_id) / to_srt() / to_vtt() / to_txt() / to_dict() + turns / speakers / num_speakers / schema_version getters. Pipeline.run_result() returns it; Pipeline.run() (plain dict) unchanged.
• .pyi stubs + py.typed marker ship in the wheel; __version__ now reads installed metadata (was a stale hardcoded string).
• to_rttm validates file_id (RTTM is whitespace-delimited — a space would silently corrupt columns).

Format writers (shared by CLI / FFI / Python)

• WebVTT transcribed cues now use the standard voice span <v SPEAKER_NN>text</v> (diarization-only cues keep the bare label).
• Cue text is sanitized for single-line payloads (newlines collapse, literal --> broken up) and VTT-escaped (& → &, < → <) — hostile/ASR text can no longer break block structure.

Docs

• examples/agent_quickstart.md — copy-paste integration paths (CLI --json, MCP, Python typed result, FFI formats, who-said-what from source).
• python README: honest collar-disclosed DER row (replaces a stale headline number) + new API listed.

Verification

• cargo nextest run --all-features: 419 passed.
• cargo clippy --all-targets --all-features -- -D warnings: clean.
• Model-gated ffi_run_format_renders_every_format (all five formats + unknown-format rejection on a valid handle) passes on real cached models.
• maturin develop + pytest: 13 passed (golden projections for all five formats, JSON round-trip, v1-consumer compat, file_id validation).
• Adversarially reviewed (two lenses + refutation pass); all confirmed findings fixed in this diff.

🤖 Generated with Claude Code

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 76.05634% with 34 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/ffi/mod.rs	67.30%	34 Missing ⚠️

📢 Thoughts on this report? Let us know!