Created attachment 99948 [details]
Diff file

Originally the HtmlFileEncodingQuery required IO every time it was used. So reading the Html file required that the 4K header of the file was red twice. For indexing 4 times as the file is read by Source used by HtmlIndexer and also TaskListIndexer. The introduced cache helped by decreasing the time in FEQ by ~ 10%. The cache caches the encoding, so it's calculated once => 3 IO ops. But the DataObject may be gced and and recreated which happens during mem intensive initial scan. So for gced DOs the 4 ops are done.
I've rewritten the htmlFEQ to work in the same way as XMLFeq. It returns a dummy charset which selects the right one when it's used and delegates to it. So there is no IO for finding the encoding. The indexing now needs only 2 IO ops per html file.

The tests are passing fine. The only change is that the returned Charset is a proxy and cannot be asked about name and so on. But it should not be a problem as most of the files including XML are doing so and no client should rely on it.

I am attaching a diff file. Marku, can you review it? If you agree I will commit it.
Thanks