feat: meeting transcription from Vexa API is stored but not chunked before embedding — transcriptions longer than the embedding model's context window are silently truncated, losing the end of long meetings

## 🐛 Problem Statement

Ell-ena's RAG pipeline (documented in `supabase/functions/generate-embeddings/`) generates vector embeddings from meeting transcriptions for semantic search. The Gemini embedding API has a **maximum input token limit** (2,048 tokens for `text-embedding-004`, or 3,072 for `embedding-001`). A meeting transcription longer than this limit is silently truncated at the embedding layer.

For a 2-hour engineering meeting (which easily produces 5,000-10,000 token transcriptions), this means the embedding only captures the first ~30 minutes. Searching for *"what did we decide about the API design in the second hour?"* returns no results — the second hour's content was never embedded.

The README documents: *"RAG (Retrieval-Augmented Generation) implementation for contextual intelligence"* — but RAG silently failing for long meetings directly contradicts this claim.

## Root Cause

`supabase/functions/generate-embeddings/index.ts` almost certainly sends the full transcription text as a single string to the embedding API without splitting it into overlapping chunks first.

## Proposed Fix

Implement a sliding window chunker before embedding generation:

```typescript
// supabase/functions/generate-embeddings/index.ts

const CHUNK_SIZE_TOKENS = 512; // Safe chunk size (well under limits)
const CHUNK_OVERLAP_TOKENS = 64; // Overlap ensures context continuity across chunks

function chunkTranscription(text: string): string[] {
// Approximate tokenization: ~4 chars per token for English
const charsPerChunk = CHUNK_SIZE_TOKENS * 4;
const overlapChars = CHUNK_OVERLAP_TOKENS * 4;
const chunks: string[] = [];

// Split on sentence boundaries to avoid cutting mid-sentence
const sentences = text.split(/(?<=[.!?])\s+/);
let currentChunk = '';

for (const sentence of sentences) {
if ((currentChunk + sentence).length > charsPerChunk) {
chunks.push(currentChunk.trim());
// Keep last `overlapChars` for continuity
currentChunk = currentChunk.slice(-overlapChars) + sentence;
} else {
currentChunk += ' ' + sentence;
}
}
if (currentChunk.trim()) chunks.push(currentChunk.trim());
return chunks;
}

// In the embedding function:
const chunks = chunkTranscription(transcriptionText);
const embeddings = await Promise.all(
chunks.map(chunk => generateEmbedding(chunk))
);

// Store each chunk as a separate embedding with metadata:
// { meeting_id, chunk_index, chunk_text, embedding, created_at }
```

Update the `meeting_vector_search` Supabase function (`sqls/09_meeting_vector_search.sql`) to return the chunk-level result but group by `meeting_id` for the final response.

## Files to Modify

| File | Change |
|---|---|
| `supabase/functions/generate-embeddings/index.ts` | Add sliding window chunker before embedding call |
| `sqls/09_meeting_vector_search.sql` | Update schema to store `chunk_index` alongside embeddings |
| `supabase/migrations/` | Add migration for `chunk_index` column |

**Suggested labels:** `bug`, `rag`, `backend`, `level: intermediate`

I would like to work on this. Could you please assign it to me?