Lightweight autocomplete for the observatory search input.

Returns up to limit suggestions across three buckets:

• tickers (4 max) — symbol + company name from PG company_profiles
• learn (2 max) — learning-path titles
• feature (2 max) — known surfaces (Markets, Bonds, Glossary, ...)

Prefix-match against in-memory caches; sub-millisecond latency. No rate limiting (the dashboard cmd-K palette has its own search at /api/search; this is the welcome-observatory's lighter cousin).