fix(translate): fix batch sizing — use real available input budget, respect job.batch_size, lazy playwright in docker

- _batch_sizer.py: compute per_batch_budget from actual available_input_budget
  (context_window - max_output_tokens) instead of sum of first N rows
- _batch_sizer.py: cap max_rows_hard_cap with job.batch_size (user config)
- _token_budget.py: add available_input_budget and max_output_tokens to return
- preview.py: validate required config fields before preview
- requirements-docker.txt: add playwright pip package (~5 MB)
- backend.entrypoint.sh: lazy playwright install chromium on first start
- build.sh: switch tar to tar.xz (-T0 -9), 5.5x smaller bundles
- README.md: add offline bundle deployment instructions
- playwright.config.js: add testMatch for *.e2e.js files
- frontend: preview tab config validation, i18n keys, translationRun store

backend/src/plugins/translate/_batch_sizer.py

@@ -20,7 +20,14 @@ from ...core.config_manager import ConfigManager
 from ...core.logger import belief_scope, logger
 from ...services.llm_provider import LLMProviderService
 from ...models.translate import TranslationJob
-from ._token_budget import PROMPT_BASE_TOKENS, estimate_token_budget
+from ._token_budget import (
+    JSON_OVERHEAD_PER_ROW,
+    MAX_OUTPUT_HEADROOM,
+    OUTPUT_PER_ROW_PER_LANG,
+    PROMPT_BASE_TOKENS,
+    REASONING_OVERHEAD,
+    estimate_token_budget,
+)
 from ._utils import estimate_row_tokens
 
 
@@ -115,15 +122,24 @@ class AdaptiveBatchSizer:
                 ]
 
             # 3. Compute per-batch row-content budget
-            estimated_input = budget.get("estimated_input_tokens", 50000)
-            per_batch_budget = estimated_input - PROMPT_BASE_TOKENS
+            # Use the ACTUAL available input capacity (context_window - max_output_tokens),
+            # NOT the sum of the first N rows. This prevents long rows from being
+            # incorrectly placed in 1-row batches when the context window has plenty of room.
+            available_input = budget.get("available_input_budget")
+            if available_input is not None:
+                # New-style: use actual available input capacity
+                per_batch_budget = available_input - PROMPT_BASE_TOKENS
+            else:
+                # Fallback for tests: use estimated_input (sum of first N rows)
+                estimated_input = budget.get("estimated_input_tokens", 50000)
+                per_batch_budget = estimated_input - PROMPT_BASE_TOKENS
 
             if per_batch_budget <= 0:
                 fallback_size = job.batch_size or 50
                 logger.explore("Per-batch budget collapsed — falling back to fixed batch size", {
-                    "estimated_input": estimated_input,
-                    "prompt_base": PROMPT_BASE_TOKENS,
+                    "per_batch_budget": per_batch_budget,
                     "fallback_size": fallback_size,
+                    "total_rows": len(source_rows),
                 })
                 return [
                     source_rows[i:i + fallback_size]
@@ -131,7 +147,26 @@ class AdaptiveBatchSizer:
                 ]
 
             # 4. Greedy batch splitting
-            max_rows_hard_cap = max(recommended * 2, 20)
+            # Compute max rows per batch from OUTPUT constraint.
+            # This prevents output truncation (finish_reason=length) when batching
+            # many short rows within the large input budget.
+            num_languages = len(target_languages)
+            max_output_tokens_val = budget.get("max_output_tokens")
+            if max_output_tokens_val is not None:
+                output_per_row = num_languages * OUTPUT_PER_ROW_PER_LANG + JSON_OVERHEAD_PER_ROW
+                available_output = max_output_tokens_val - REASONING_OVERHEAD - MAX_OUTPUT_HEADROOM
+                max_rows_by_output = max(available_output // output_per_row, 1) if output_per_row > 0 else 20
+                max_rows_hard_cap = max_rows_by_output
+            else:
+                # Fallback for tests: old formula
+                max_rows_hard_cap = max(recommended * 2, 20)
+
+            # 5. Respect job.batch_size as the absolute maximum rows per batch.
+            #    User-configured batch_size overrides model-based estimates to
+            #    prevent LLM quality degradation on large batches.
+            if job.batch_size:
+                max_rows_hard_cap = min(max_rows_hard_cap, job.batch_size)
+
             batches: list[list[dict]] = []
             current_batch: list[dict] = []
             current_tokens = 0

backend/src/plugins/translate/_token_budget.py

@@ -235,6 +235,8 @@ def estimate_token_budget(
             estimated_output_tokens: Total estimated output tokens.
             max_output_needed: Recommended max_tokens for this batch.
             warning: str | None if batch was reduced.
+            available_input_budget: Total input token budget per batch (context_window - max_output_tokens).
+            max_output_tokens: The effective max output tokens limit used for this batch.
     """
     if not target_languages:
         target_languages = ["en"]
@@ -333,6 +335,8 @@ def estimate_token_budget(
         "estimated_output_tokens": estimated_output,
         "max_output_needed": max_output_needed,
         "warning": warning,
+        "available_input_budget": available_input_budget,
+        "max_output_tokens": max_output_tokens,
     }
 # endregion estimate_token_budget
 # #endregion estimate_token_budget

backend/src/plugins/translate/preview.py

@@ -80,6 +80,10 @@ class TranslationPreview:
                 raise ValueError("Job must have a source datasource configured for preview")
             if not job.translation_column:
                 raise ValueError("Job must have a translation column configured for preview")
+            if not job.target_languages:
+                raise ValueError("Job must have at least one target language configured for preview")
+            if not job.provider_id:
+                raise ValueError("Job must have an LLM provider configured for preview")
 
             target_languages = job.target_languages or [job.target_dialect or "en"]
             if not isinstance(target_languages, list):