fix(futuresearch-mcp): defer polling/results instructions to per-client tool responses (#5066)

## Summary

- The MCP system instructions unconditionally told the agent to use
`futuresearch_progress` for polling and `futuresearch_results` for
output retrieval
- On widget-capable clients (Claude Desktop, Claude.ai), this
contradicted the per-call tool responses that direct the agent to
`futuresearch_status` instead, causing the agent to ignore the widget
and double-poll with the wrong tool
- Updated workflow steps 3 (Poll) and 4 (Results), plus the HTTP Results
section, to defer to the submission/polling tool responses which already
branch correctly per client

## Test plan

- [ ] Verify on Claude Desktop: after submitting an operation, agent
calls `futuresearch_status` and does NOT follow up with
`futuresearch_progress` or `futuresearch_results`
- [ ] Verify on Claude Code: agent still uses `futuresearch_progress` →
`futuresearch_results` flow as before
- [ ] Verify on everyrow-cc: same text-only flow as Claude Code

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Sourced from commit 118b92bc9b06c67a41a52239fe445cb39f73c16b

Author: RafaelPo

futuresearch-mcp/src/futuresearch_mcp/app.py

@@ -124,9 +124,12 @@ async def no_auth_http_lifespan(_server: FastMCP):
 2. **Submit** — call a processing tool (futuresearch_agent, futuresearch_classify, \
 futuresearch_rank, futuresearch_dedupe, futuresearch_merge, futuresearch_forecast). \
 It returns a task_id immediately.
-3. **Poll** — call `futuresearch_progress(task_id)` repeatedly until the task completes. \
-Do NOT add commentary between progress calls — just call again immediately.
-4. **Results** — call `futuresearch_results(task_id)` to retrieve the output.
+3. **Poll** — follow the instruction in the submission response. It will tell you \
+exactly which tool to call next (`futuresearch_status` or `futuresearch_progress`). \
+Do NOT override that instruction — use the tool it specifies.
+4. **Results** — follow the instruction from the polling tool. If it says results \
+are in the widget, do NOT call futuresearch_results. Otherwise, call \
+`futuresearch_results(task_id)` to retrieve the output.
 
 ## Session and artifact reuse
 
@@ -174,7 +177,9 @@ def _build_instructions_http() -> str:
 - Do NOT pass local file paths to `futuresearch_upload_data` — it will fail in remote mode.
 
 ## Results
-- Always call `futuresearch_results(task_id, page_size=N)` immediately when the task completes. \
+- Follow the instruction from the polling tool. If it says results are in the widget, \
+do NOT call futuresearch_results — the widget handles display automatically.
+- Otherwise, call `futuresearch_results(task_id, page_size=N)` when the task completes. \
 For small tasks ({threshold} or fewer rows), set page_size to the total. \
 For larger tasks, set page_size to {threshold} to load the first batch.
 - The user always has access to all rows via the table view and download link.