feat(rag): notify browser of LLM retry/waiting/fallback/warning and search hit count

[marevol]

Summary

Surfaces previously-invisible in-progress events from the AI Search (RAG/Chat) flow to the browser UI via new SSE event types and an extended phase-complete payload.

New SSE events

Event	Payload	When
retry	{phase, operation, attempt, maxAttempts, sleepMs, cause?}	LLM HTTP call retry (fired once per retry attempt by the LLM plugin)
waiting	{phase, reason, elapsedMs, timeoutMs}	When streamChatWithConcurrencyControl blocks on an exhausted permit
fallback	{phase, reason, originalQuery, newQuery}	Before re-running search with a regenerated query (no_results / no_relevant_results)
warning	{phase, code, detail}	When intent detection silently falls back due to reasoning-model token exhaustion

The existing event: phase (status: complete) payload is now extended for the search phase to include hitCount so the browser can show "N documents found" before the longer evaluate/fetch phases finish.

Changes

• LlmStreamCallback — adds default no-op onRetry / onWaiting / onWarning. @FunctionalInterface preserved.
• ChatPhaseCallback — adds default no-op onRetry / onWaiting / onFallback / onWarning and a payload-aware onPhaseComplete(String, Map<String,Object>) that defaults to delegating to the legacy single-arg form.
• PhaseAwareStreamCallback (new) — bridges LLM-layer LlmStreamCallback events to phase-aware ChatPhaseCallback events.
• ChatClient.streamChatEnhanced — wraps every LlmStreamCallback lambda with PhaseAwareStreamCallback, emits onFallback before query regeneration, includes hitCount in search phase completion, fires onWarning when intent detection falls back.
• AbstractLlmClient.streamChatWithConcurrencyControl — fires onWaiting when the concurrency permit is unavailable.
• IntentDetectionResult — adds isFallback() flag (additive; only fallbackSearch(...) sets true).
• ChatApiManager — emits the new SSE events. New emitSseEventSafely + putIfNotNull helpers reduce boilerplate.
• chat.js / chat.jsp — adds listeners for the new events and i18n labels.
• fess_label*.properties — adds 6 new keys across all 17 language files (English source for non-translated languages, Japanese translations included).

Backwards compatibility

All interface additions are default no-op methods. Existing implementers (including the 3 LLM plugins) compile and behave unchanged. Plugin PRs to actually fire onRetry follow this PR — see "Companion PRs" below.

Companion PRs (must merge AFTER this one and a new SNAPSHOT is published)

• codelibs/fess-llm-openai — invoke LlmStreamCallback#onRetry from executeWithRetry
• codelibs/fess-llm-ollama — same
• codelibs/fess-llm-gemini — same

Until those merge, the retry SSE event will not fire (rest of the events fire from this PR alone).

Test plan

• mvn test — 8 touched test classes, 72 tests, 0 failures
  • LlmStreamCallbackTest (1)
  • ChatPhaseCallbackTest (9 — 7 preexisting preserved + 2 new)
  • PhaseAwareStreamCallbackTest (4)
  • AbstractLlmClientWaitingTest (2)
  • ChatClientFallbackTest (1)
  • ChatClientHitCountTest (2)
  • ChatClientWarningTest (2)
  • ChatApiManagerTest (51 — 47 preexisting preserved + 4 new)
• mvn formatter:format && mvn license:format — clean (no-op)
• node -c chat.js — JS syntax OK