the-claw: Add performance benchmarks for ingestion, SQLite, and embed paths

[clean6378-max-it]

Planned Effort

5 story points — sprint item #7 (Medium): Performance benchmarks for ingestion, SQLite, and LLM-related paths

One issue → one PR. claude-code-chat-browser export_count work is a separate issue/PR (Thursday sprint #4).

Depends on

Prerequisite	Why
the-claw repo cloned	Work lives in the-claw/claw
pnpm install	Vitest + native deps (better-sqlite3)
Unit tests green on base branch	pnpm --filter @the-claw/claw test before/after

No dependency on the browser export_count PR (different repo). Can run in parallel after morning browser work or on another machine.

Problem

No performance baselines exist for ingestion throughput, SQLite read paths, or LLM/embed latency. Regressions in the webhook → queue → store pipeline are invisible until production feels slow.

Goal

Add a local, repeatable benchmark harness with committed baseline numbers and a short README. No blocking CI gate this week (optional informational workflow step is fine).

Scope

1. Directory and docs

• claw/benchmarks/README.md — prerequisites, run command, scenario descriptions, how to refresh baselines, CI note (optional/non-blocking).

2. Vitest bench (or tsx timing scripts)

Scenario	Suggested target
Ingestion throughput	src/pipeline/queue.ts — IngestionQueue push/ack (in-memory or temp SQLite)
SQLite read latency	src/warehouse/messages.ts (or query helpers) — temp DB + seed rows
LLM / embed path	Mocked embed/LLM client — measure orchestration only (no live API keys in CI)

Suggested files:

• claw/benchmarks/ingestion.bench.ts
• claw/benchmarks/sqlite-read.bench.ts
• claw/benchmarks/embed-path.bench.ts

3. Committed baselines

• claw/benchmarks/baselines.json (or baselines.md) with date, environment note, and key metrics.

4. Package script

Add to claw/package.json:

"bench": "vitest bench"

Document run: pnpm --filter @the-claw/claw bench

Acceptance Criteria

• Benchmark suite runnable locally (vitest bench documented in README)
• Covers ingestion, SQLite read, and LLM/embed-related timing (network mocked)
• Baseline numbers committed under claw/benchmarks/
• README explains interpretation and refresh process
• pnpm --filter @the-claw/claw test still green (benchmarks separate from unit tests)