RAG-2: BM25 keyword index over scraped Live docs

[nonogv]

Context

Depends on #66 (RAG-1). Part of the v2 RAG milestone.

What

Implement a lightweight BM25 scorer (~80 lines, no external deps) over data/live-docs.json.

File: src/agent/docs-index.ts

Export:
```ts
/** Returns the top-k most relevant chunks for a query. Returns [] if index is absent. */
export function retrieveLiveDocs(query: string, topK?: number): DocChunk[];

/** True if the local docs index is older than maxAgeDays (default 30). */
export function isDocIndexStale(maxAgeDays?: number): boolean;
```

Implementation notes:

• Tokenise query and corpus: lowercase, split on word boundaries, strip punctuation
• BM25 with standard defaults: k1=1.5, b=0.75
• Return top topK (default 3) chunks above a minimum relevance threshold
• If data/live-docs.json is absent, return [] silently — never throw

Tests in src/agent/docs-index.test.ts:

• Returns correct top result for known device queries ("Operator FM", "Glue Compressor ratio")
• Returns [] when data file is absent
• Respects topK limit
• isDocIndexStale returns true for an index older than the threshold

Acceptance criteria

• retrieveLiveDocs('Operator FM synthesis') returns a chunk mentioning Operator from the manual
• retrieveLiveDocs('anything') returns [] when data/live-docs.json does not exist
• No external runtime dependencies added
• All existing tests still pass