Fix sheet-extraction page loss: bare-list wrap, compact retry, reasoning cap

- SheetExtractorAgent accepts top-level array responses as the objects
  array instead of discarding them (recovered the failure mode behind
  16/38 failed sheets on job 475a6f184dd1)
- Second-chance compact retry per page before declaring extraction failed
- call_json: reasoning_effort param (cloud-only extra_body), finish_reason
  capture + explicit max_tokens log line, finish_reason in raw dumps,
  cache key covers reasoning_effort
- EXTRACT_MAX_TOKENS default 16384 -> 32768 (Gemini thinking tokens count
  against the cap); new EXTRACT_REASONING_EFFORT=low default for extract
- tests: 5 new fallback-ladder tests

backend/config.py

@@ -73,7 +73,14 @@ PDF_DPI       = int(os.getenv("PDF_DPI",       "100"))
 MAX_PAGES     = int(os.getenv("MAX_PAGES",     "60"))
 MAX_DIMENSION = int(os.getenv("MAX_DIMENSION", "2400"))  # px cap on the long edge
 LLM_TIMEOUT   = int(os.getenv("LLM_TIMEOUT",   "180"))   # seconds per call
-EXTRACT_MAX_TOKENS = int(os.getenv("EXTRACT_MAX_TOKENS", "16384"))
+# Gemini 2.5 Pro counts thinking tokens against max_tokens, so the visible
+# JSON budget is well under this number on dense sheets. 32768 leaves real
+# headroom; raise via env if a set still truncates.
+EXTRACT_MAX_TOKENS = int(os.getenv("EXTRACT_MAX_TOKENS", "32768"))
+# Reasoning effort for the per-sheet extractor (OpenRouter reasoning knob).
+# Extraction is perceptive, not deliberative - "low" keeps thinking tokens
+# from eating the output budget. Empty string disables the parameter.
+EXTRACT_REASONING_EFFORT = os.getenv("EXTRACT_REASONING_EFFORT", "low").strip()
 REASON_MAX_TOKENS  = int(os.getenv("REASON_MAX_TOKENS",  "4096"))
 
 # -- QAQC stage knobs (Stages 0-1, 3, 6-11) -------------------------