test: integration + unit tests for cache fix and OUTPUT_SAFETY_FACTOR

## TestClassifyCacheSourceLang (8 tests) — regression for source-lang exclusion
- test_source_lang_in_tls_cache_all_translations: exact prod bug scenario
  (detected_lang=ru, tls=[ru,en,fr,zh], cache={en,fr,zh} → pre_rows not LLM)
- test_source_lang_in_tls_cache_missing_one_translation: partial cache → LLM
- test_detected_und_tls_includes_und: und excluded correctly
- test_empty_detected_lang_no_exclusion: '' → no filtering
- test_source_lang_not_in_tls: no-op exclusion when lang not in tls
- test_multi_row_mixed_source_lang_cache: 4 rows, correct split
- test_all_same_lang_short_circuit: single target lang shortcut
- test_case_insensitive_detected_lang_matching: RU→ru

## TestBatchPipelineE2E (3 tests) — full pipeline integration
- test_warm_cache_full_pipeline_no_llm: 17 rows all cached → 0 LLM
- test_cold_cache_pipeline_all_llm: no cache → all LLM
- test_partial_cache_pipeline_split: 5 cached + 5 not → correct split

## TestOutputSafetyFactor (4 tests) — OUTPUT_SAFETY_FACTOR invariants
- test_output_safety_factor_not_above_070: ≤0.70 guard
- test_max_rows_by_output_qwen_flash_4langs: 14-24 range
- test_output_safety_factor_consistent_with_per_row: manual = actual
- test_single_lang_output_rows_above_20: ≥20 rows for 1 lang

## Semantic fixes
- Add missing #endregion TestClassifyCacheSourceLang
- Move floating @BRIEF tags inside their regions (REASONING_OVERHEAD, DICT_TOKENS_PER_ENTRY)
- Add @BRIEF to OUTPUT_PER_ROW_PER_LANG region

Verification: 705 tests pass (0 failures, 2 pre-existing warnings)

backend/pyproject.toml

@@ -17,7 +17,6 @@ include = ["src*"]
 [tool.pytest.ini_options]
 pythonpath = ["."]
 asyncio_mode = "auto"
-testpaths = ["tests", "src/plugins"]
 markers = [
     "integration: Integration tests requiring external services (Docker, Testcontainers, Superset). Use --run-integration to enable.",
 ]

backend/src/plugins/translate/__tests__/test_batch_classify_persist.py

@@ -394,6 +394,9 @@ class TestClassifyCacheSourceLang:
     # #endregion test_case_insensitive_detected_lang_matching
 
 
+# #endregion TestClassifyCacheSourceLang
+
+
 # ── _persist_pre() Tests ───────────────────────────────────────────
 
 

backend/src/plugins/translate/_token_budget.py

@@ -24,9 +24,9 @@ DEFAULT_CONTEXT_WINDOW = 64000
 
 DEFAULT_MAX_OUTPUT_TOKENS = 16384
 # #endregion DEFAULT_MAX_OUTPUT_TOKENS
-# @BRIEF CoT reasoning overhead tokens for DeepSeek models (~2000 tokens for chain-of-thought).
 # #region REASONING_OVERHEAD [TYPE Constant]
 # @ingroup Translate
+# @BRIEF CoT reasoning overhead tokens for DeepSeek models (~2000 tokens for chain-of-thought).
 
 REASONING_OVERHEAD = 2000
 # #endregion REASONING_OVERHEAD
@@ -53,6 +53,7 @@ PROVIDER_DEFAULTS: dict[str, dict[str, int]] = {
 # #endregion PROVIDER_DEFAULTS
 # #region OUTPUT_PER_ROW_PER_LANG [TYPE Constant]
 # @ingroup Translate
+# @BRIEF Output tokens budgeted per row per target language for batch sizing.
 
 #            Increased from 120 to 200 because production logs show finish_reason=length
 #            even with 120 tok/row/lang — long source texts produce long translations.
@@ -71,9 +72,9 @@ JSON_OVERHEAD_PER_ROW = 50
 #            Increased from 300 to 600 to account for longer template, system msg, and dict preamble.
 PROMPT_BASE_TOKENS = 600
 # #endregion PROMPT_BASE_TOKENS
-# @BRIEF Cap for dictionary tokens in estimation to avoid overestimating.
 # #region DICT_TOKENS_PER_ENTRY [TYPE Constant]
 # @ingroup Translate
+# @BRIEF Cap for dictionary tokens in estimation to avoid overestimating.
 
 DICT_TOKENS_PER_ENTRY = 20
 # #endregion DICT_TOKENS_PER_ENTRY