RAG-1: Live docs scraper script

[nonogv]

Context

Part of the v2 RAG milestone. See BACKLOG.md for the full dependency chain.

What

Add scripts/scrape-live-docs.ts, runnable as npm run scrape-docs.

Crawls and chunks the following into data/live-docs.json:

Source	Content
ableton.com/en/manual/	Full Ableton Live manual — instrument/audio effect/MIDI effect reference pages
makingmusic.ableton.com	Intermediate music-making concepts (server-rendered, fetchable)
ableton.com/en/blog/	Selected tutorials (tag-filtered: tutorials, sound-design)

Each chunk:
```ts
interface DocChunk {
id: string; // stable hash of url + offset
url: string;
title: string; // page title or section heading
body: string; // ~400 token text excerpt
tags: string[]; // e.g. ['device', 'Operator', 'synthesis']
source: 'manual' | 'makingmusic' | 'blog';
scrapedAt: string; // ISO date
}
```

Output: data/live-docs.json (target ≤ 5MB) and data/live-docs-diff.json (added/removed/updated URLs + unchanged count).

The scraper must be idempotent — re-running updates changed pages, keeps unchanged ones.

Notes

• learningmusic.ableton.com is a JS SPA; static fetch returns a shell. Skip it.
• Ableton's manual uses anchor sections (e.g. #operator) on three reference pages, not per-device URLs. The abledtonManualUrl() function in web-search.ts already maps these correctly — the scraper should use the same anchor extraction logic (extractDeviceSectionFromHtml).
• Scraper exits non-zero on hard failures so CI can catch structural changes.

Acceptance criteria

• npm run scrape-docs produces data/live-docs.json with ≥ 50 chunks
• Re-running is idempotent (same content = same chunk ids, no churn)
• data/live-docs-diff.json is written with added/removed/updated/unchanged counts
• Hard failures (network outage, 404 on all three reference pages) exit non-zero