Fixes three performance bottlenecks identified in code review of
recently merged features (PR #121 conversations viewer, PR #190 AI feedback).

## Changes

### migrations/004_performance_indexes.cjs (new)
Four new indexes targeting hot query paths:

- idx_ai_feedback_guild_created (guild_id, created_at DESC)
  getFeedbackTrend() and getRecentFeedback() filtered by guild_id
  AND created_at but only had a single-column guild_id index, forcing
  a full guild scan + sort on every trend/recent call.

- idx_conversations_content_trgm (GIN, pg_trgm)
  content ILIKE '%...%' search was a sequential scan. GIN/trgm index
  reduces this from O(n) to O(log n * trigram matches).
  Requires pg_trgm extension (added idempotently).

- idx_conversations_guild_created (guild_id, created_at DESC)
  Default 30-day listing query filters guild_id + created_at. The
  existing 3-column (guild_id, channel_id, created_at) composite is
  suboptimal when channel_id is not in the predicate.

- idx_flagged_messages_guild_message (guild_id, message_id)
  Conversation detail + flag endpoints query flagged_messages by
  guild_id AND message_id = ANY(...). Existing index only covers
  (guild_id, status).

### src/api/routes/conversations.js
**GET / — Replace in-memory pagination with SQL CTE grouping**

Before: fetched up to 10,000 message rows into Node memory, grouped
them in JavaScript (O(n) time + memory), then sliced for pagination.
Every page request loaded the full 10k row dataset.

After: single SQL query using window functions (LAG + SUM OVER) to
identify conversation boundaries and aggregate summaries directly.
COUNT(*) OVER() provides total count without a second query.
Pagination happens at the DB with LIMIT/OFFSET on summary rows.
Memory overhead is now proportional to page size (default 25), not
total conversation volume.

Removed now-unused buildConversationSummary() helper (logic inlined
into the SQL-side aggregation).

**POST /:conversationId/flag — Parallel verification queries**

Before: msgCheck and anchorCheck ran sequentially (~2× RTT).
After: both run in parallel via Promise.all (1× RTT for verification).

### tests/api/routes/conversations.test.js
Updated 'should return paginated conversations' test to mock the new
SQL CTE response shape (pre-aggregated summary rows) instead of raw
message rows. All 41 conversation tests pass.

- Add ORDER BY to getTemplateByName for deterministic results
- Fix roleId precedence to preserve existing roleIds during merge
- Truncate Discord embed field values to 1024 chars

- Filter out options with empty roleIds before saving
- Only enable role menu for non-built-in templates with valid options
- Add user-facing note when options are filtered

- validateTemplateOptions now validates that optional roleId and
  description fields are strings when present
- Update JSDoc @see reference from issue #135 (voice tracking) to
  issue #216 (role menu templates)
- Update ON CONFLICT clause to use constraint name for consistency
  with the new LOWER(name) index