perf(embeddings): cross-node batching + worker pool

[theagenticguy]

Summary

• Refactors the embeddings phase from one-embedding-per-node-per-await into two stages: a job-collection pass that walks symbol/file/community tiers in canonical order producing {text, emitRow} records, and a dispatch loop that fires workers × batchSize embeds concurrently per wave and scatters vectors back into the row buffer.
• Adds a Piscina pool of independent OnnxEmbedder workers (packages/ingestion/src/pipeline/phases/embedder-{worker,pool}.ts). Each worker holds its own ONNX session; the pool is exposed behind an Embedder-shaped facade so the phase doesn't branch. A main-thread canary OnnxEmbedder opens first so EmbedderNotSetupError keeps its class identity across the structured-clone boundary.
• New flags: --embeddings-workers <n|auto> and --embeddings-batch-size <n> (defaults: 1 and 32 — unchanged single-threaded behaviour out of the box).

Motivation

Real-world codehub analyze --embeddings --force --granularity symbol,file,community on a ~1,922-file AWS codebase sat at 95% CPU for 7+ minutes before the refactor. The phase was awaiting embedBatch() per node inside a single-threaded ONNX session (intraOpNumThreads: 1, graphOptimizationLevel: "disabled" — required for the graphHash determinism contract), so there was no concurrency anywhere in the stack.

Determinism

The graphHash / embeddingsHash contract is preserved:

• Canonical tier ordering (symbol → file → community) is unchanged.
• Rows are still sorted by (granularity, nodeId, chunkIndex) before hashing.
• openOnnxEmbedder()'s deterministic knobs are intact per worker — which input produces which vector is independent of which worker ran it.
• New regression test asserts embeddingsHash at batchSize=1 equals embeddingsHash at batchSize=32.

Expected speedup

On an M-series laptop with --embeddings-workers auto --embeddings-batch-size 32, the 7-minute AWSQuickWork run should drop to roughly 1–2 minutes. --embeddings-int8 cuts that further.

Test plan

• pnpm build — clean
• pnpm --filter @opencodehub/ingestion test — 576/576 pass
• New test: embeddings.test.ts — batchSize=1 vs batchSize=32 produce byte-identical embeddingsHash
• codehub analyze --help surfaces --embeddings-workers and --embeddings-batch-size
• End-to-end: run codehub analyze AWSQuickWork --embeddings --force --granularity symbol,file,community --embeddings-workers auto and confirm wall time drop + identical embeddingsHash vs a single-threaded control run