docs(15.9): add the operating facts of the chunk job, embedding settings and engine-side fusion

[marevol]

Summary

Add to the 15.9 semantic search pages the operating facts they left out, and correct one statement they got wrong. Each statement was re-read in the Fess source at the current main (66b1fa64b) before it was written.

All seven languages (ja, en, de, es, fr, ko, zh-cn), development tree (15.9) only. The ja and en text is the source; de, es, fr, ko and zh-cn are translated from it, reusing each page's existing terms.

Changes

• config/rank-fusion.rst
  • Hit counts: with engine-side fusion the total hit count and the facet counts cover the union of the keyword and semantic hits (each searcher contributes at most rank.fusion.pagination_depth results per shard, the semantic searcher at most content_chunker.search.knn.k). With Fess-side fusion the count is the main searcher's count plus the existing correction (at most window_size / 2), and the facets are the main searcher's, so the two modes report different counts for the same query.
  • Paging: a page whose start is rank.fusion.pagination_depth or more is refused; in the v2 search API this is HTTP 400 invalid_request.
  • rank.fusion.timeout does not apply to engine-side fusion, so an unresponsive embedding provider is bounded by its own timeout times its attempts (built-in opensearch provider: timeout x retry.max, 60 s x 3 by default, plus the waits between attempts).
• config/search-semantic.rst
  • content_chunker.embedding.dimension: an unset value gives a 768-dimension mapping without a warning (the page said it warned; only an empty, non-numeric, non-positive or over-16000 value warns). The mapping's dimension cannot be changed afterwards. New section If the Index Was Created Without a Dimension: what happens (the runtime read has no fallback, so documents become fail) and how to recover.
  • content_chunk_status: skipped covers empty or whitespace-only bodies, documents with no chunk produced, documents over max_chunks_per_document, and every document when content_chunker.chunker.name names no chunker. skipped and fail are terminal (fail is selected again only with content_chunker.job.retry_failed=true). How to read the status distribution as the state of the chunk job, and how to clear skipped after fixing the cause.
  • content_chunker.embedding.opensearch.retry.max is the maximum number of attempts including the first (3 = three requests, two waits), not the number of retries.
  • content_chunker.search.knn.k is per shard; with Fess-side fusion the effective value has a floor of rank.fusion.window_size divided by the number of searchers (100 by default), so a smaller value has no effect; with engine-side fusion it is used as configured.
  • chunk_size is in characters while the model's limit is in tokens, and Fess does not truncate the text it embeds; a measured example for paraphrase-multilingual-MiniLM-L12-v2.
  • New section Changing to Another Model with the Same Dimension: it is accepted silently because only the dimension is checked, old vectors and new query vectors then live in different spaces, and how to switch.
• api/api-search.rst: q is limited by api.param.max.length (default 1000); a longer value returns 400 invalid_request. This is not a search engine limit.

config/properties.rst is generated and is not touched.

Not documented

Behaviour of how the chunk job handles provider failures and what its summary line reports is deliberately left out of the new text.

Verification

• python3 -m unittest discover -s tools -p 'test_*.py': 79 tests OK
• python3 tools/update_eol.py --check, python3 tools/gen_properties_doc.py --check, python3 tools/check_headings.py on the 15.8 and 15.9 trees: pass
• Sphinx build (default builder, minimal configuration) of each language's 15.9 config/ and api/ directories before and after the change: the set of warnings is identical, and no unrendered inline markup remains in the changed pages
• docutils parse (doctitle_xform=False) of every changed file: no new messages, and the two new sections, the new notes and the new literal blocks are present in each language