Since Lucene 2.9 we have the method getSequentialSubReader() returning IndexReader. Based on hardly-to-debug errors in user's code, Robert and me realized that returning an array from a public API is an anti-pattern. If the array is intended to be modifiable (like byte in terms,...), it is fine to use arrays in public APIs, but not, if the array must be protected from modification. As IndexReaders are 100% unmodifiable in trunk code (no deletions,...), the only possibility to corrumpt the reader is by modifying the array returned by getSequentialSubReaders(). We should prevent this.
The same theoretically applies to FieldCache, too - but the party that is afraid of performance problems is too big to fight against that
For getSequentialSubReaders there is no performance problem at all. The binary search of reader-ids inside BaseCompositeReader can still be done with the internal protected array, but public APIs should expose only a unmodifiable List. The same applies to leaves() and children() in IndexReaderContext. This change to list would also allow to make CompositeReader and CompositeReaderContext Iterable<R extends IndexReader>, so some loops would look nice.