v1.4.0

Response-handling fixes (mirrored with the remote MCP server — identical behavior):

• Never silently empty: a search reporting finished before its results are written now keeps polling through a settle window and returns status: "results_pending" (pointing at get_keyword_search) instead of finished with results: null.
• Response size cap (~100 KB): large responses no longer break MCP transports. The largest post arrays are trimmed from the end with an explicit truncated: true + truncation report; analytics/sentiment/summary fields are always complete. Full post data remains available via the posts tools with a platform filter. Configurable via ROLLI_MAX_RESPONSE_BYTES.
• Boolean query guidance: keyword_search documents the backend's boolean syntax (AND / OR / NOT, parentheses, quoted exact phrases) so compound queries stop token-OR-matching into unrelated posts.
• Credit preflight: start_agent_run / rerun_agent_run check the IQ credit balance first and fail instantly with required_credits vs available_credits — no more runs failing mid-flight on credits.
• Balance visibility: get_usage now includes credits.credits_remaining (plus credits_limit / period_reset_date once the backend ships them).
• Agent runs now poll inline for 2 minutes then return the run ID for async retrieval (matching the remote server); report_html is stripped from run payloads.
• Platforms aligned with the remote server: threads added; linkedin and weibo removed.
• New env vars: ROLLI_MAX_RESPONSE_BYTES (default 100000), ROLLI_RESULTS_SETTLE_MS (default 45000).