Keep merge inside --optimize

[rejojer]

Merge now belongs to --optimize rather than running on every build, so the
three levels each mean something distinct:

	tree
default	the extracted tree, as-is
--optimize merge	deterministic merge, no LLM
--optimize	merge, then the LLM expansion pass

Previously the default collapsed the tree too, which left --optimize merge
with almost nothing extra to do and made the search-cost metrics credit
--optimize with a gain the default already delivered.

The tradeoff is that the default now summarises every extracted node: about 16%
more summary calls across the example corpus, and up to 48% on a document whose
outline has many one-page subsections. Callers who want the cheaper, flatter
tree ask for it with --optimize merge, which costs nothing beyond the merge
itself.

key_items only appears once a merge has run, so it is now documented as an
--optimize field. The non-Flash pipeline in page_index.py is unchanged: it
has no optimize switch and has always collapsed, so it keeps doing so.