fix(coverage): admit protected-base test retirement

CHANGELOG.md

@@ -45,6 +45,13 @@ Semantic Versioning where the repository publishes a release.
 
 ### Fixed
 
+- Keep the post-merge replay guard focused on lost protected-base evidence: a
+  feature-only source and its same-subject test may be removed together when
+  both are absent from the exact protected base, while protected-base test
+  deletion, test-only deletion, weakened retained tests, exact tree replay,
+  targeted base unmerge, and bulk regression remain fail-closed. Declare pip as
+  a development dependency because the materialized-lock regression invokes
+  its hash preflight inside the project-local uv environment.
 - Used the receiving repository's workflow token for same-repository scheduler
   Actions inventory and read calls, while retaining the established mutation
   credential chain. An exhausted organization-wide OpenCode App installation

docs/doctoring/pr-head-replay-base-alignment.md

@@ -0,0 +1,70 @@
+# PR-head replay guard: protected-base alignment
+
+## Observed failure
+
+LineageWeave pull request 258 exact head
+`6dc040c6b3ea0bfc4424bb7afb11b8afd7205d77` failed central OpenCode coverage
+run `32528869351`. The replay guard found that five commits after the newest
+protected-base-descended merge removed three files, including
+`tests/test_lineage_contract.py`.
+
+The exact protected base
+`ef6f5a5ffcb467bd935dc1e53acc0029669b0bd7` already omitted both that test and
+its same-subject runtime module. The head also omitted both. No path had been
+restored to pre-merge content, the head did not replay an earlier tree, and the
+three-file delta stayed below the conservative bulk-replay threshold. The
+test-regression signal nevertheless treated every post-merge test deletion as
+protected-base evidence loss.
+
+## Decision and security boundary
+
+The guard now distinguishes one base-alignment case from an evidence
+regression. A test absent from the exact protected base is not reported as
+regressed only when a non-test file with the same logical subject and language
+suffix is deleted in the same post-merge range and that source is also absent
+from the protected base, with no same-subject source remaining in the exact
+head. This admits removal of a feature-only source/test pair that aligns the
+head with the protected branch without letting a same-named documentation or
+sibling-source deletion excuse a code-test deletion.
+
+The following cases remain blocking:
+
+- a test or same-subject source present on the protected base is removed;
+- a feature-only test is removed while its source remains;
+- a retained test file loses declared test cases without a replacement test
+  file;
+- the head exactly replays a pre-merge tree, restores a path to its pre-merge
+  content, or crosses the conservative bulk-deletion thresholds.
+
+This boundary follows NIST SSDF's root-cause and verification practices while
+preserving GitHub's guidance that privileged pull-request workflows must treat
+pull-request state as untrusted input. The decision is based only on immutable
+Git objects from the validated base, merge anchor, and head; no pull-request
+code executes during classification.
+
+## Verification
+
+- A new Git-history regression reproduces a feature-only source/test pair,
+  merges the exact protected base, removes the pair, and requires a passing
+  replay decision.
+- Existing regressions continue to require failure for protected-base source
+  and test removal, test-only deletion, weakened retained tests, exact replay,
+  targeted unmerge, and bulk deletion.
+- The patched guard passes against the observed LineageWeave base/head pair and
+  reports no regressed protected-base test path.
+- The focused replay-guard suite, complete central Python quality suite,
+  docstring coverage, workflow validation, and diff hygiene run on the final
+  tree.
+- The project-local uv environment declares pip because the existing bounded
+  include regression invokes `python -m pip` for its hash preflight; verification
+  no longer depends on an unrecorded manual venv seed.
+
+## References
+
+GitHub. (n.d.). *Secure use reference*. GitHub Docs. Retrieved August 22, 2026,
+from https://docs.github.com/en/actions/reference/security/secure-use
+
+Souppaya, M., Scarfone, K., & Dodson, D. (2022). *Secure software development
+framework (SSDF) version 1.1: Recommendations for mitigating the risk of
+software vulnerabilities* (NIST Special Publication 800-218). National
+Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.800-218

pyproject.toml

@@ -6,6 +6,7 @@ dependencies = []
 
 [dependency-groups]
 dev = [
+    "pip>=25.0",
     "pytest>=8.0.0",
     "pytest-cov>=7.1.0",
     "interrogate>=1.7.0",

scripts/ci/pr_head_replay_guard.py

@@ -16,8 +16,10 @@
   merge brought in (observed in appguardrail#297, where a stale snapshot
   reverted an accessibility wrapper and deleted its regression tests in a
   push far below the bulk thresholds);
-- test regression without replacement: post-merge commits deleting test files
-  or reducing declared test cases while adding no replacement test file;
+- test regression without replacement or protected-base alignment: post-merge
+  commits deleting test files or reducing declared test cases while adding no
+  replacement test file, except when a feature-only test and its same-subject,
+  same-language source are both retired to match the protected base;
 - the conservative bulk-regression signature: at least five tracked files and
   500 lines removed, with deletions at least four times additions.
 
@@ -91,7 +93,7 @@ def unmerges_base_work(self) -> bool:
 
     @property
     def suspicious_test_regression(self) -> bool:
-        """Return whether test cases were lost with no replacement test file."""
+        """Return whether non-base-aligned tests were lost without replacement."""
         return bool(self.regressed_test_paths) and self.added_test_files == 0
 
     @property
@@ -188,6 +190,20 @@ def is_test_path(path: str) -> bool:
     )
 
 
+def logical_test_subject(path: str) -> str:
+    """Return the source-like stem represented by a common test filename."""
+    name = Path(path).name
+    for marker in (".test.", ".spec."):
+        if marker in name:
+            return name.split(marker, 1)[0]
+    stem = Path(name).stem
+    if stem.startswith("test_"):
+        return stem.removeprefix("test_")
+    if stem.endswith("_test"):
+        return stem.removesuffix("_test")
+    return stem
+
+
 def changed_paths(repo_root: Path, start: str, end: str) -> set[str]:
     """Return the set of paths whose content differs between two commits."""
     output = git_output(repo_root, ["diff", "--name-only", start, end])
@@ -234,16 +250,64 @@ def test_case_count(
     return len(pattern.findall(source)) if pattern is not None else None
 
 
-def test_file_changes(repo_root: Path, start: str, end: str) -> tuple[tuple[str, ...], int]:
-    """Return deleted or test-case-reducing paths and the added-test count."""
+def test_file_changes(
+    repo_root: Path,
+    start: str,
+    end: str,
+    protected_revision: str,
+) -> tuple[tuple[str, ...], int]:
+    """Return non-base-aligned test regressions and the added-test count."""
+    protected_paths = set(
+        git_output(
+            repo_root,
+            ["ls-tree", "-r", "--name-only", protected_revision],
+        ).splitlines()
+    )
+    head_source_subjects: set[tuple[str, str]] = set()
+    for path in git_output(
+        repo_root,
+        ["ls-tree", "-r", "--name-only", end],
+    ).splitlines():
+        if not is_test_path(path):
+            head_source_subjects.add(
+                (logical_test_subject(path), Path(path).suffix.lower())
+            )
     regressed: set[str] = set()
     added = 0
-    for line in git_output(repo_root, ["diff", "--name-status", start, end]).splitlines():
+    name_status_lines = git_output(
+        repo_root,
+        ["diff", "--name-status", start, end],
+    ).splitlines()
+    deleted_paths: set[str] = set()
+    removed_feature_sources: set[tuple[str, str]] = set()
+    for line in name_status_lines:
+        fields = line.split("\t")
+        if len(fields) < 2 or not fields[0].startswith("D"):
+            continue
+        deleted_paths.add(fields[-1])
+        source_subject = (
+            logical_test_subject(fields[-1]),
+            Path(fields[-1]).suffix.lower(),
+        )
+        if (
+            not is_test_path(fields[-1])
+            and fields[-1] not in protected_paths
+            and source_subject not in head_source_subjects
+        ):
+            removed_feature_sources.add(source_subject)
+    for line in name_status_lines:
         fields = line.split("\t")
         if len(fields) < 2 or not is_test_path(fields[-1]):
             continue
         status = fields[0][:1]
-        if status == "D":
+        if status == "D" and (
+            fields[-1] in protected_paths
+            or (
+                logical_test_subject(fields[-1]),
+                Path(fields[-1]).suffix.lower(),
+            )
+            not in removed_feature_sources
+        ):
             regressed.add(fields[-1])
         elif status == "A":
             added += 1
@@ -252,7 +316,11 @@ def test_file_changes(repo_root: Path, start: str, end: str) -> tuple[tuple[str,
         if len(fields) < 3 or not fields[0].isdigit() or not fields[1].isdigit():
             continue
         path = fields[2]
-        if not is_test_path(path) or int(fields[1]) <= int(fields[0]):
+        if (
+            not is_test_path(path)
+            or path in deleted_paths
+            or int(fields[1]) <= int(fields[0])
+        ):
             continue
         before_count = test_case_count(repo_root, start, path)
         after_count = test_case_count(repo_root, end, path)
@@ -293,7 +361,12 @@ def collect_evidence(repo_root: Path, base_sha: str, head_sha: str) -> ReplayEvi
         head_sha,
     )
     unmerged = unmerged_base_paths(repo_root, merge_anchor, head_sha)
-    regressed_tests, added_tests = test_file_changes(repo_root, merge_anchor, head_sha)
+    regressed_tests, added_tests = test_file_changes(
+        repo_root,
+        merge_anchor,
+        head_sha,
+        base_sha,
+    )
     return ReplayEvidence(
         base_sha=base_sha,
         head_sha=head_sha,
@@ -336,7 +409,8 @@ def format_report(evidence: ReplayEvidence) -> str:
         )
     if evidence.suspicious_test_regression:
         reasons.append(
-            "post-merge commits deleted test files or reduced declared test cases "
+            "post-merge commits deleted protected-base or unpaired feature test files, "
+            "or reduced declared test cases "
             "without adding any replacement test file: "
             f"{summarize_paths(evidence.regressed_test_paths)}."
         )

tests/test_pr_head_replay_guard.py

@@ -256,6 +256,73 @@ def test_test_refactor_with_replacement_passes(tmp_path):
     assert not evidence.blocked
 
 
+def test_removing_feature_only_test_to_match_current_base_passes(tmp_path):
+    """Deleting a test absent from the protected base is not a base replay."""
+    repo = tmp_path / "repo"
+    repo.mkdir()
+    git(repo, "init", "-b", "main")
+    git(repo, "config", "user.name", "Test")
+    git(repo, "config", "user.email", "test@example.com")
+    write(repo, "README.md", "base\n")
+    commit(repo, "base")
+
+    git(repo, "checkout", "-b", "feature")
+    write(repo, "src/legacy.py", "def legacy():\n    return True\n")
+    write(repo, "tests/test_legacy.py", "def test_legacy():\n    assert True\n")
+    commit(repo, "feature adds legacy code")
+
+    git(repo, "checkout", "main")
+    write(repo, "base-change.txt", "current base\n")
+    current_base = commit(repo, "advance base without legacy code")
+
+    git(repo, "checkout", "feature")
+    git(repo, "merge", "--no-ff", "--no-edit", "main")
+    (repo / "src/legacy.py").unlink()
+    (repo / "tests/test_legacy.py").unlink()
+    git(repo, "add", "-A")
+    head = commit(repo, "align feature with protected base")
+
+    evidence = guard.collect_evidence(repo, current_base, head)
+
+    assert evidence.regressed_test_paths == ()
+    assert evidence.unmerged_paths == ()
+    assert not evidence.blocked
+
+
+def test_removing_same_named_sibling_source_does_not_excuse_test_loss(tmp_path):
+    """A retained same-subject source keeps its feature test protected."""
+    repo = tmp_path / "repo"
+    repo.mkdir()
+    git(repo, "init", "-b", "main")
+    git(repo, "config", "user.name", "Test")
+    git(repo, "config", "user.email", "test@example.com")
+    write(repo, "README.md", "base\n")
+    commit(repo, "base")
+
+    git(repo, "checkout", "-b", "feature")
+    write(repo, "src/primary/legacy.py", "def legacy():\n    return True\n")
+    write(repo, "src/obsolete/legacy.py", "def legacy():\n    return False\n")
+    write(repo, "tests/test_legacy.py", "def test_legacy():\n    assert True\n")
+    commit(repo, "feature adds same-subject sources")
+
+    git(repo, "checkout", "main")
+    write(repo, "base-change.txt", "current base\n")
+    current_base = commit(repo, "advance base")
+
+    git(repo, "checkout", "feature")
+    git(repo, "merge", "--no-ff", "--no-edit", "main")
+    (repo / "src/obsolete/legacy.py").unlink()
+    (repo / "tests/test_legacy.py").unlink()
+    git(repo, "add", "-A")
+    head = commit(repo, "remove sibling and test")
+
+    evidence = guard.collect_evidence(repo, current_base, head)
+
+    assert evidence.regressed_test_paths == ("tests/test_legacy.py",)
+    assert evidence.suspicious_test_regression
+    assert evidence.blocked
+
+
 def test_is_test_path_covers_common_layouts():
     """Test-path detection recognizes directories, prefixes, suffixes, and spec names."""
     assert guard.is_test_path("tests/test_guard.py")
@@ -267,6 +334,11 @@ def test_is_test_path_covers_common_layouts():
     assert guard.is_test_path("tests\\test_windows.py")
     assert not guard.is_test_path("scripts/ci/guard.py")
     assert not guard.is_test_path("docs/testing.md")
+    assert guard.logical_test_subject("tests/test_guard.py") == "guard"
+    assert guard.logical_test_subject("pkg/guard_test.go") == "guard"
+    assert guard.logical_test_subject("app/button.test.tsx") == "button"
+    assert guard.logical_test_subject("app/button.spec.ts") == "button"
+    assert guard.logical_test_subject("tests/README.md") == "README"
 
 
 def test_test_case_count_fails_closed_and_supports_known_formats(
@@ -343,6 +415,7 @@ def test_test_file_changes_parses_status_and_numstat(monkeypatch, tmp_path):
             "D\ttests/test_gone.py",
             "A\ttests/test_new.py",
             "M\ttests/test_kept.py",
+            "D\tdocs/gone.md",
             "D\tsrc/app_old.py",
             "badline",
         ]
@@ -355,15 +428,22 @@ def test_test_file_changes_parses_status_and_numstat(monkeypatch, tmp_path):
             "2\t9\tsrc/big.py",
         ]
     )
-    outputs = iter([name_status, numstat])
+    outputs = iter(
+        [
+            "tests/test_shrunk.py",
+            "src/kept.py",
+            name_status,
+            numstat,
+        ]
+    )
     monkeypatch.setattr(guard, "git_output", lambda _root, _args: next(outputs))
     monkeypatch.setattr(
         guard,
         "test_case_count",
         lambda _root, revision, _path: 2 if revision == "a" else 1,
     )
 
-    regressed, added = guard.test_file_changes(tmp_path, "a", "b")
+    regressed, added = guard.test_file_changes(tmp_path, "a", "b", "protected")
 
     assert regressed == ("tests/test_gone.py", "tests/test_shrunk.py")
     assert added == 1