improve tooling and whitespace error handling

inference.py

@@ -96,6 +96,13 @@ Action policy:
 - After RUN_TESTS, do not choose RUN_TESTS again immediately unless test evidence is genuinely missing.
 - Treat "no output" as invalid reasoning when pass_count_summary or traceback text is present.
 
+Strict state transition rules (no looping on same action):
+- If pass_count_summary=unknown (no test output yet), MUST run RUN_TESTS next (never VIEW_CODE or REPLACE_LINES).
+- After VIEW_CODE, prefer RUN_TESTS next to get test evidence (never do VIEW_CODE twice in a row).
+- After REPLACE_LINES, always do RUN_TESTS next to validate the fix (never do REPLACE_LINES twice in a row).
+- If REPLACE_LINES+RUN_TESTS did not fix the issue, do VIEW_CODE next to re-orient before another edit.
+- Do not attempt the same edit twice; if it fails, change strategy (VIEW_CODE or UNDO_EDIT or RESET_TO_ORIGINAL).
+
 Worked examples (generic, no benchmark task leakage):
 
 Example 1: failing tests after RUN_TESTS -> choose REPLACE_LINES
@@ -128,6 +135,29 @@ Observation: syntax_error=true and traceback provides a concrete syntax failure
 Valid action JSON:
 {"thought":"Observation: ... Diagnosis: ... Plan: ...","action_type":"REPLACE_LINES","start_line":8,"end_line":9,"new_code_block":"    # syntax-fixed code"}
 
+Example 4: NO TEST OUTPUT YET + UNKNOWN PASS COUNT -> MUST choose RUN_TESTS (not VIEW_CODE)
+Input evidence snippet:
+- pass_count_summary=unknown
+- all_tests_pass_signal=false
+- last_execution_output is empty ""
+Valid thought:
+Observation: no test execution has occurred; pass_count_summary is unknown and no test output is available. Diagnosis: cannot diagnose code bugs without seeing test failures; must run tests first. Plan: choose RUN_TESTS to collect test evidence and determine what needs fixing.
+Valid action JSON (WRONG - violates state rule):
+BAD: {"thought":"...","action_type":"VIEW_CODE","start_line":null,"end_line":null,"new_code_block":null}
+CORRECT: {"thought":"...","action_type":"RUN_TESTS","start_line":null,"end_line":null,"new_code_block":null}
+
+Example 5: EDIT FAILED - SAME ACTION TWICE IS WRONG
+Previous step: REPLACE_LINES on line 4 (indentation fix attempt 1)
+Current observation:
+- syntax_error=true (indentation still wrong)
+- same error message about dedentation
+- pass_count_summary=unknown
+Valid thought:
+Observation: same indentation error persists despite the previous REPLACE_LINES attempt. The fix did not work. Diagnosis: attempting the identical REPLACE_LINES again will create an infinite loop; must change strategy. Plan: call VIEW_CODE next to re-examine the context and understand why the fix didn't apply properly, then revise the approach.
+Valid action JSON (WRONG - causes looping):
+BAD (same edit): {"thought":"...","action_type":"REPLACE_LINES","start_line":4,"end_line":4,"new_code_block":"        fixed code"}
+CORRECT (change strategy): {"thought":"...","action_type":"VIEW_CODE","start_line":null,"end_line":null,"new_code_block":null}
+
 Submit gate (hard rule):
 - If any failure, error, traceback, xfailed/unfinished signal, or uncertainty remains, do not SUBMIT.
 - If all-tests-passed signal is present, do SUBMIT immediately on this turn.
@@ -436,13 +466,10 @@ async def run(difficulty: Optional[str] = None, show_thought: bool = False) -> N
                 consecutive_same_action_count = 1
                 last_action_type = current_action_type
 
-            if (
-                current_action_type == "RUN_TESTS"
-                and consecutive_same_action_count >= 3
-            ):
+            if consecutive_same_action_count >= 3:
                 kill_switch_triggered = True
                 history.append(
-                    "KILL_SWITCH: RUN_TESTS selected 3 times consecutively. "
+                    f"KILL_SWITCH: {current_action_type} selected 3 times consecutively. "
                     "Terminating episode early to prevent looping."
                 )
                 steps_taken = step