docs(llm): correct the 15.8 LLM overview against the implementation

[marevol]

Summary

Reviewed 15.8/config/llm-overview.rst against the implementation and corrected it, then propagated the result to all seven languages (ja, en, de, es, fr, ko, zh-cn).

Sources of truth: fess (master) AbstractLlmClient / LlmClientManager / ChatClient / ChatPhaseCallback / fess_config.properties, and the fess-llm-ollama, fess-llm-openai and fess-llm-gemini plugins.

Corrections

Topic	Was	Now
availability.check.interval	"0 disables it"	The value is read via getConfigInt(), which accepts only values > 0, so 0 falls back to the default 60. The check cannot be disabled through this property.
System prompt customization	"You must edit the DI XML inside the JAR"	Documents the supported LastaDi component redefinition file app/WEB-INF/classes/fess_llm+<componentName>.xml, with the provider-to-component-name table.
rag.llm.name	"Only the Ollama client is enabled by default"	The default is ollama, and the value selects the DI component name {rag.llm.name}LlmClient. Leaving the default while installing only a non-Ollama plugin enables no client at all and logs [LLM] LlmClient not found. componentName=ollamaLlmClient. Also documents none.
fess_config.properties path	app/WEB-INF/conf/	app/WEB-INF/classes/ (and /etc/fess/ for package installs), consistent with the rest of the docs.
Provider-specific parameters	Implied all providers support thinking.budget, top.p and reasoning.effort	Per-provider support matrix. reasoning.effort is OpenAI only, top.p is not available for Gemini, and thinking.budget is ignored by OpenAI. Also notes that the default.* fallback tier is Ollama only.
Architecture	One unconditional eight-step flow	Scoped to the streaming search intent. Query regeneration is a search-phase fallback, not a phase, and the non-streaming POST /api/v2/chat runs no evaluation phase.
OpenAI description	GPT-4	GPT-5, matching the gpt-5-mini default.

Two .. warning:: blocks were added to the system prompt section: a redefinition file replaces the whole component definition (omitted properties revert to unset), and copying fess_llm++.xml itself into app/WEB-INF/classes/ registers the component twice and prevents startup with TooManyRegistrationComponentException.

Additions

• rag.chat.enabled default value (false).
• Plugin installation through the admin UI, in addition to manual JAR placement.
• A note that the direct prompt type is not invoked in the current version.
• Links to the chat search guide and the chat API reference.

Verification

• The system prompt override behaviour was confirmed empirically against lasta-di 2.0.0, not only by reading the source: a redefinition file overrides a component declared in a plugin's fess_llm++.xml, while a second fess_llm++.xml on the classpath fails with TooManyRegistrationComponentException.
• All seven files build without warnings under Sphinx.
• All seven files are structurally identical: 44 table rows, 114 literal tokens and 12 :doc: targets each. The only literal differences between languages are two placeholders that are prose rather than identifiers (fess_llm+{component name}.xml and rag.llm.ollama.default.{parameter}).
• Values confirmed unchanged because they are already correct: the three default models and endpoints, the authentication headers, rag.chat.* defaults, max.concurrent.requests (5), concurrency.wait.timeout (30000, and the resulting error really is a rate-limit error), chat.evaluation.max.relevant.docs (3), the ten prompt type names, and the five phase names.

Three Korean section underlines that were shorter than their East Asian display width are also fixed; they produced Sphinx warnings and predate this change.

Follow-ups (not in this PR)

• rag-chat.rst documents copying fess_llm++.xml into app/WEB-INF/, which is the procedure that breaks startup, and describes the concurrency wait timeout as a timeout error when the code raises a rate-limit error.
• llm-ollama.rst, llm-openai.rst, llm-gemini.rst and rag-chat.rst all still use the app/WEB-INF/conf/fess_config.properties path.
• The 15.7 copies of this page carry the same claims, but should be checked against the 15.7 branch rather than master.