Here's a number of theoretical considerations.
Index-based access is fine for selecting very small numbers of rows, but as the number of rows required rises then fine-grained partitioning + table scans with partition pruning become a much more viable option, and as you say you are limited in your choice of partitioning options with IOT's.
Furthermore the IOT is forced into a particular physical row order, and that may not be sympathetic to the order that would most benefit queries (through reducing clustering factor on common access paths) or data compression. It also requires more resources to maintain the structure.
Speaking of compression, with a heap table you can get compression based on all columns of the table, not just the PK columns (which I think is the restriction you'd face on an IOT).
I expect that access to the overflow columns would also be a significant performance hit (not tested it though).
I'd run some very careful tests on the relative performance of accessing the metrics through bitmap indexes -> mapping table -> IOT table -> overflow in comparison to bitmap indexes -> heap fact table.
Fact tables are supposed to be nothing but foreign keys to dimension tables plus "facts" - numeric measurements.
From the 10g Concepts manual, on the benefits of IOT -
"Index-organized tables provide faster access to table rows by the primary key or any key that is a valid prefix of the primary key. Presence of nonkey columns of a row in the B-tree leaf block itself avoids an additional block access. Also, because rows are stored in primary key order, range access by the primary key (or a valid prefix) involves minimum block accesses.
In order to allow even faster access to frequently accessed columns, you can use a row overflow segment (as described later) to push out infrequently accessed nonkey columns from the B-tree leaf block to an optional (heap-organized) overflow segment. This allows limiting the size and content of the portion of a row that is actually stored in the B-tree leaf block, which may lead to a higher number of rows in each leaf block and a smaller B-tree.
Unlike a configuration of heap-organized table with a primary key index where primary key columns are stored both in the table and in the index, there is no such duplication here because primary key column values are stored only in the B-tree index.
Because rows are stored in primary key order, a significant amount of additional storage space savings can be obtained through the use of key compression.
Use of primary-key based logical rowids, as opposed to physical rowids, in secondary indexes on index-organized tables allows high availability. This is because, due to the logical nature of the rowids, secondary indexes do not become unusable even after a table reorganization operation that causes movement of the base table rows. At the same time, through the use of physical guess in the logical rowid, it is possible to get secondary index based index-organized table access performance that is comparable to performance for secondary index based access to an ordinary table."
Since the fact tables will be the start point for the majority of queries I want the access to the result set to be as fast as possible. A current complaint is the length of time that it takes to get a result. I plan to move most of the measurements to the overflow area and make the IOT portion just the foreign keys to the dimension tables (the fact table's primary key). That's what led me to the idea of a fact table as an IOT. The fact table consists of it's primary key plus some measurements. Since I use an index to enforce that primary key I have a situation where most of the data that I want from the fact table can be satisfied from the index alone. I decided that I might as well save the space of the table (since the table and the index will basically duplicate each other's data) and make the table index organized. I'm partitioning it by date since most queries are limited to a single fiscal year.
But I need a mechanism to speed queries that are not via the primary key or a valid prefix thereof. That's where the secondary indexes come in.
I'm a little dissapointed that IOT tables cannot be composite partitioned. My dimension tables will be range partitioned by date (because of the date limited nature of the queries) and hash subpartitioned to spread the I/O over as many spindles as possilbe (each subpartition goes to its own tablespace and each tablespace will be on a separate spindle).
However, that's just my thinking at the planning stage. If you have experience with the flaws of using IOT fact tables I certainly want to learn from that experience and not repeat a known mistake.
What are the problems with IOT fact tables?