test(agents/plan_generation): fix BDD step ambiguity and add tdd_bug tag (#10480)

The compliance tests added in 80fa1329 had three issues that broke
multiple CI gates:

1. unit_tests (Behave): the new
   features/steps/validation_bypass_issue_10480_steps.py redefined
   `@given("the plan generation coverage module is imported")`, which
   was already defined in plan_generation_coverage_boost_steps.py.
   Behave raised AmbiguousStep and 32 scenarios errored on import.
   The duplicate has been removed; the feature reuses the existing
   step via behave's global step discovery.

2. unit_tests (Behave): three `@then` decorators included a literal
   "and " prefix (e.g. `@then("and no artificial length threshold ...")`),
   which never matches because Behave normalises "And" into the
   preceding step's keyword and strips it from the match text. The
   prefixes have been removed so the three previously-undefined steps
   now resolve.

3. tdd_quality_gate: scenarios lacked the `@tdd_bug_10480` tag that
   scripts/tdd_quality_gate.py requires to recognise the regression
   guard for the bug. All 5 scenarios now carry the tag.

Also fixes a grammatical error in the CHANGELOG entry
("responsible for determine" → "responsible for determining") noted
by the reviewer.

Refs #10480

CHANGELOG.md

@@ -7,7 +7,7 @@
 - Fixed a validation bypass in `PlanGenerationGraph._validate()` (issue #10480):
   the artificial ``or len(all_code) > 10`` condition that caused any code longer
   than 10 characters to auto-pass LLM validation has been removed.  LLM verdicts
-  are now solely responsible for determine pass/fail status, enabling the retry
+  are now solely responsible for determining pass/fail status, enabling the retry
   logic and broken-code rejection to operate correctly.
 
 - Hardened the TDD bug-fix quality gate for issue #629: PR parsing now

features/steps/validation_bypass_issue_10480_steps.py

@@ -40,10 +40,10 @@ def make_fake_llm(responses: list[str]) -> FakeListLLM:
 # ---------------------------------------------------------------------------
 # Background steps
 # ---------------------------------------------------------------------------
-@given("the plan generation coverage module is imported")
-def step_coverage_module_imported(context):
-    """Ensure the plan_generation module is importable."""
-    assert PlanGenerationGraph is not None
+# Note: "the plan generation coverage module is imported" is defined in
+# features/steps/plan_generation_coverage_boost_steps.py and reused here
+# (behave discovers steps globally across features/steps/). Defining it
+# again would raise AmbiguousStep.
 
 
 @given("the validation bypass test helper is ready")
@@ -168,7 +168,7 @@ def step_validation_message_contains_feedback(context):
     )
 
 
-@then("and no artificial length threshold bypassed the rejection")
+@then("no artificial length threshold bypassed the rejection")
 def step_no_length_threshold_bypass(context):
     """Confirm that code length did not override a FAIL verdict.
 
@@ -185,7 +185,7 @@ def step_no_length_threshold_bypass(context):
     )
 
 
-@then("and the validation message should preserve the full LLM feedback")
+@then("the validation message should preserve the full LLM feedback")
 def step_validation_preserves_feedback(context):
     """Full message must contain the original LLM response."""
     assert context.validation_result is not None
@@ -196,7 +196,7 @@ def step_validation_preserves_feedback(context):
     )
 
 
-@then("and the code-length condition must NOT override the FAIL verdict")
+@then("the code-length condition must NOT override the FAIL verdict")
 def step_code_length_not_override(context):
     """Definitive regression test: code length > 10 did not cause auto-pass.
 

features/validation_bypass_issue_10480.feature

@@ -7,30 +7,35 @@ Feature: Validation bypass fix (issue #10480)
     Given the plan generation coverage module is imported
     And the validation bypass test helper is ready
 
+  @tdd_bug_10480
   Scenario: LLM returns FAIL for short code and validation correctly rejects it
     Given an LLM mock that responds with "FAIL: syntax error"
     When a PlanGenerationGraph _validate is called with short generated code
     Then the validation result status should be "FAIL"
     And the validation message should contain the LLM feedback
 
+  @tdd_bug_10480
   Scenario: LLM returns FAIL for long code and validation correctly rejects it
     Given an LLM mock that responds with "FAIL: unsafe pattern detected"
     When a PlanGenerationGraph _validate is called with longer generated code
     Then the validation result status should be "FAIL"
     And no artificial length threshold bypassed the rejection
 
+  @tdd_bug_10480
   Scenario: LLM returns PASS and validation correctly accepts it
     Given an LLM mock that responds with "PASS"
     When a PlanGenerationGraph _validate is called with any generated code
     Then the validation result status should be "PASS"
     And the validation message should contain the LLM feedback
 
+  @tdd_bug_10480
   Scenario: LLM returns FAIL with additional details and validation correctly rejects
     Given an LLM mock that responds with "FAIL: missing error handling in auth module"
     When a PlanGenerationGraph _validate is called with generated code
     Then the validation result status should be "FAIL"
     And the validation message should preserve the full LLM feedback
 
+  @tdd_bug_10480
   Scenario: Validation ignores code length and does not auto-pass long code
     Given an LLM mock that responds with "FAIL: security vulnerability"
     When a PlanGenerationGraph _validate is called with code longer than 10 characters