fix(mcp): escalate transient failures before asking for a restart (#6353)

* fix(mcp): escalate transient failures before asking for a restart

Give dropped connections and startup failures one bounded retry, then a
scripted handoff to the user. Point timeout recovery at a named request
field or REA_* setting instead of telling agents those limits are fixed.

Co-authored-by: morluto <morluto@users.noreply.github.com>

* fix(process): stop unexpected command failures from recommending a retry

The catch can already have launched the target, and the projection stays
non-retryable. Tell the user to check the inputs and run doctor instead.

Co-authored-by: morluto <morluto@users.noreply.github.com>

* fix(mcp): limit retries to safe calls and recommended settings

A dropped connection may already have changed the target, and a bad
catalog rejects the same startup every time. Retry only a repeatable
call, and change a setting only when the remediation recommends it.

Co-authored-by: morluto <morluto@users.noreply.github.com>

* test(mcp): match recovery guidance across wrapped lines

Startup handoff capitalization and markdown wrapping still carry the
same sentences. Compare the prose with whitespace collapsed.

Co-authored-by: morluto <morluto@users.noreply.github.com>

---------

Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: morluto <morluto@users.noreply.github.com>

docs/mcp-contracts.md

@@ -95,11 +95,15 @@ optional bounded `recovery` next step. `recovery.kind=inspect_input_contract`
 names the rejected operation: re-read that operation's advertised input schema
 (call `tools/list` if the catalog may be stale), correct the listed fields, and
 do not repeat the unchanged call. `recovery.kind=inspect_timeout_state` names
-the operation and the REA server/provider deadline; inspect partial Evidence
-and reported cleanup state, then follow producer-specific remediation or reduce
-supported work. Generic errors do not assume adjustable timeout controls. Raising
-only the MCP client's deadline cannot extend this budget. These hints never
-include the complete catalog or caller arguments.
+the operation and the REA server/provider deadline. Inspect partial Evidence
+and reported cleanup state, then follow the returned remediation. Change a
+request field or REA_* setting only when the returned remediation explicitly
+recommends that change. Restart REA when that text says the value is read at
+startup. A mention that a control does not raise the deadline is not a
+recommendation to change it. When it names no setting, reduce supported work;
+do not invent one. Raising only the MCP
+client's deadline cannot extend this budget. These hints never include the
+complete catalog or caller arguments.
 Failures omit `structuredContent`, since each advertised `outputSchema`
 describes successful results. Successful replies retain their schema-validated
 structured data and matching text projection. Oversized errors are retained