feat(pipeline_v2): optional dense embedding (sub-window per segment)

[ekhodzitsky]

What

Adds opt-in dense embedding for pipeline v2: PipelineConfig.embed_window_secs: Option<f32> (CLI --v2 --embed-window <secs>, bench --embed-window). Some(w) slides a w-second window (hop w/2) inside each primary segment so a speaker run produces several embeddings — like the legacy pipeline's dense windows — instead of one. More embeddings per speaker → more robust centroids → lower confusion. Pure compute, no extra model weights.

Why — measured (collar 0.25, macro DER)

The triangle (legacy / v2+AHC / v2+VBx) showed legacy wins clean VoxConverse largely because of its dense sliding-window embeddings (confusion 5.9 vs v2's 12.1). This closes that gap for v2.

set	v2+AHC sparse	v2+AHC + dense(1.5)
VoxConverse-30	16.4% (conf 12.1)	14.7% (conf 9.7)
VoxConverse-10 (harder)	19.8%	15.7%
AMI-4 (overlap)	32.8%	32.5% (beats legacy 35.5%)

w = 1.5 is the sweet spot — 1.0 over-fragments (18.7%, worse than sparse).

Honest scope

• Real v2 improvement (VoxConverse −1.7pp, harder subsets −4pp, beats legacy on AMI).
• Does not overtake legacy on clean VoxConverse (13.5%), and slightly hurts the VBx clusterer on AMI (VBx-sparse stays the meeting champion).
• So it ships opt-in (None default = current behavior, one embedding per segment). It's a v2/AHC booster, not a default change.
• Cost: more embedder calls (≈ legacy's density), so slower than sparse v2.

Tests

• expand_embed_units unit tests (splits long segments, keeps short, None = identity, sub-windows inherit the local speaker index and end at the segment boundary).
• CLI --help snapshot updated.
• e2e verified via bench (--embed-window 1.5).
• fmt clean · clippy --all-targets clean · 310 lib tests green.

Breaking (v2 API)

PipelineConfig gained embed_window_secs — struct-literal constructors add embed_window_secs: None (or use ..default()). Covered by the 0.9.0 minor bump.

🤖 Generated with Claude Code

[codecov-commenter]

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

Codecov Report

❌ Patch coverage is 92.30769% with 5 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/bin/polyvoice.rs	25.00%	3 Missing ⚠️
src/bin/polyvoice-bench.rs	0.00%	1 Missing ⚠️
src/pipeline_v2/mod.rs	98.30%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!