[SWE-Paddle] PaddlePaddle__77749 fix

swe-paddle/tasks/PaddlePaddle__Paddle-77749/README.md

@@ -25,19 +25,19 @@ The task adds `paddle.nn.utils.rnn.pad_sequence` and `paddle.nn.utils.rnn.unpad_
 ## Test Classification
 
 - **F2P:** all 21 cases in the new `test/legacy_test/test_rnn_utils.py` file. They cannot import the new module on the exact base and pass after the solution.
-- **P2P:** the exact gold test patch adds no independent P2P cases because the entire target file is new. Existing Paddle utility and Tensor behavior remains an external regression expectation.
+- **P2P:** `test/legacy_test/test_rnn_cell_api.py::TestRnnUtil::test_case`, an untouched RNN utility regression guard for `paddle.utils.map_structure` / `paddle.utils.assert_same_structure`. The gold diff does not modify that file, and the node passes on both the exact base and gold revisions.
 
 ## Artifact Map
 
 - `proposal.md`: approved source proposal; preserved unchanged.
 - `instruction.md`: self-contained observable requirements for the public APIs.
 - `solution/code.patch`: exact base-to-gold diff for the two production files.
 - `tests/test.patch`: exact base-to-gold diff adding the unittest file.
-- `tests/test.sh`: strict direct-Python unittest wrapper.
+- `tests/test.sh`: wrapper that runs the existing RNN utility P2P node before the new F2P API test file.
 - `environment/README.md`: revisions, runtime assumptions, patch order, and verification guidance.
 
 ## Verification Summary
 
 - Both patches are exported from the exact base-to-gold Git diff with the requested production/test boundary.
 - Packaging checks cover patch whitespace, shell syntax, isolated test-first application order, byte-for-byte gold equivalence, and Python syntax.
-- Runtime execution is performed with isolated exact-source overlays when the exact historical native build is unavailable; the runtime commit and results are recorded in `environment/README.md`.
+- The selected `TestRnnUtil` P2P node is byte-for-byte unchanged between the base and gold revisions and passed repeated runs on both isolated overlays; the base target failed at the `paddle.nn.utils.rnn` import and the gold target ran all 21 cases with `OK`. Runtime details are recorded in `environment/README.md`.

swe-paddle/tasks/PaddlePaddle__Paddle-77749/environment/README.md

@@ -28,11 +28,19 @@ git apply /path/to/PaddlePaddle__Paddle-77749/solution/code.patch
 PYTHON_BIN=python bash /path/to/PaddlePaddle__Paddle-77749/tests/test.sh
 ```
 
-The first run is expected to fail while importing the absent `paddle.nn.utils.rnn` module. After applying the solution and refreshing the runtime's Python package, all 21 target cases should pass.
+The first run must show the existing P2P RNN utility node passing, followed by the expected failure while the target file imports the absent `paddle.nn.utils.rnn` module. After applying the solution and refreshing the runtime's Python package, both the P2P node and all 21 target cases should pass.
 
-The wrapper deliberately executes the legacy unittest file directly:
+The wrapper runs the P2P guard and then the F2P target:
 
 ```bash
+# P2P guard: existing RNN utility behavior
+(
+    cd test/legacy_test
+    "${PYTHON_BIN:-python}" -m pytest -q \
+        test_rnn_cell_api.py::TestRnnUtil::test_case
+)
+
+# F2P target: newly added pad_sequence / unpad_sequence tests
 "${PYTHON_BIN:-python}" test/legacy_test/test_rnn_utils.py
 ```
 
@@ -41,11 +49,13 @@ The wrapper deliberately executes the legacy unittest file directly:
 - The available loadable runtime is 363 commits newer than the exact base. It supplies best-effort native support but is not an exact historical runtime.
 - A pre-solution overlay that leaves the newer runtime's RNN utility module in place invalidates F2P verification.
 - These APIs rely on Tensor creation, concatenation, stacking, transposition, and slicing behavior. A substantially different runtime can introduce unrelated failures.
-- The exact test patch contains only new F2P coverage, so broad P2P confidence requires existing Paddle regression suites beyond this target wrapper.
+- The exact test patch contains only new F2P coverage, so the wrapper adds the existing RNN utility node `test/legacy_test/test_rnn_cell_api.py::TestRnnUtil::test_case` as an explicit P2P guard. Broader P2P confidence still relies on the rest of Paddle's regression suites.
 - Do not patch, reset, clean, or otherwise alter the dirty `/workspace/Paddle` checkout. Use an isolated snapshot or worktree.
 
 ## Local Verification Record
 
-Package validation passed for artifact presence, unchanged `proposal.md`, shell syntax, patch whitespace, exact file boundaries, test-first then solution application, byte-for-byte equality with the gold revision, and Python syntax compilation. Neither the wrapper nor the environment commands invokes an alternate test runner.
+Package validation passed for artifact presence, unchanged `proposal.md`, shell syntax, patch whitespace, exact file boundaries, test-first then solution application, byte-for-byte equality with the gold revision, and Python syntax compilation. The wrapper invokes `pytest` only for the pre-existing P2P node; the new target file is still executed directly.
 
-Runtime validation used isolated overlays backed by the loadable runtime at `ae907b878e91dbabf3582da99f8b05a46b588fc2`, 363 commits after the exact base. The base overlay replaced the utility exports with the exact base file and removed the newer runtime's RNN utility module; the test-only run then failed during import with `ModuleNotFoundError`, as expected. The gold overlay used both exact gold production files, and the complete direct-Python wrapper ran 21 tests with `OK`. Because the exact test patch is an entirely new file, this target run provides F2P coverage only and does not claim an independent P2P result. The dirty `/workspace/Paddle` checkout was not modified.
+Runtime validation used isolated overlays backed by the loadable runtime at `ae907b878e91dbabf3582da99f8b05a46b588fc2`, 363 commits after the exact base. The base overlay replaced the utility exports with the exact base file and removed the newer runtime's RNN utility module; the target run then failed during import with `ModuleNotFoundError`, as expected. The gold overlay used both exact gold production files, and the complete direct-Python target ran 21 tests with `OK`. The dirty `/workspace/Paddle` checkout was not modified.
+
+After adding the P2P guard, an additional stability check used isolated exact base/gold overlays with an installed Paddle `3.5.0.dev20260814` CPU runtime on macOS. `test/legacy_test/test_rnn_cell_api.py::TestRnnUtil::test_case` passed 5/5 on the base overlay and 5/5 on the gold overlay; the base target still failed with `ModuleNotFoundError`, and the gold target ran 21 tests with `OK`.

swe-paddle/tasks/PaddlePaddle__Paddle-77749/tests/test.sh

@@ -3,4 +3,14 @@ set -euo pipefail
 
 PYTHON_BIN="${PYTHON_BIN:-python}"
 
+# P2P (pass-to-pass): existing RNN utility regression guard. The gold patch
+# does not modify test/legacy_test/test_rnn_cell_api.py, so this node must pass
+# both before and after the solution is applied.
+(
+    cd test/legacy_test
+    "${PYTHON_BIN}" -m pytest -q \
+        test_rnn_cell_api.py::TestRnnUtil::test_case
+)
+
+# F2P (fail-to-pass): all cases in the newly added API test file.
 "${PYTHON_BIN}" test/legacy_test/test_rnn_utils.py