lusoris · May 28, 2026 · May 28, 2026
@@ -0,0 +1,13 @@
+# BREAKING: Remove VmafLegacyQualityRunner (float-path runner)
+
+`VmafLegacyQualityRunner` has been removed from the Python harness.
+The class depended on the `float_ansnr` feature extractor, which was
+dropped from the C backend in PR #38. Any code importing or calling
+`VmafLegacyQualityRunner` will receive an `ImportError`.
+
+**Migration**: use `VmafQualityRunner` with a current VMAF model
+(e.g. `vmaf_v0.6.1.json`). The modern runner uses the integer-path
+feature extractors and is numerically equivalent to the legacy runner
+for all content types.
+
+ADR-0749.
@@ -460,15 +460,11 @@ class VmafIntegerFeatureExtractor(VmafFeatureExtractor):
             [f"integer_{f}" for f in VmafFeatureExtractor.ATOM_FEATURES],
         )
     )
-<<<<<<< HEAD
     # float_ansnr was removed in feat(core): drop legacy ansnr feature (#38).
     # The "ansnr" → "float_ansnr" mapping is intentionally omitted here so
     # requests via VmafIntegerFeatureExtractor never pass the removed extractor
     # name to the CLI.  The float/legacy path (VmafFeatureExtractor) retains
     # the mapping for backward-compatibility of the legacy test suite.
-=======
-    ATOM_FEATURES_TO_VMAFEXEC_KEY_DICT["ansnr"] = "float_ansnr"
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
     ATOM_FEATURES_TO_VMAFEXEC_KEY_DICT["anpsnr"] = "float_anpsnr"
 
     def _generate_result(self, asset):
@@ -483,14 +479,10 @@ def _generate_result(self, asset):
         h = quality_height
         logger = self.logger
 
-<<<<<<< HEAD
         # float_ansnr was removed in feat(core): drop legacy ansnr feature (#38).
         # The VmafQualityRunner (integer path) only needs adm/vif/motion to
         # produce VMAF scores; ansnr was never a model input for vmaf_v0.6.1.
         features = ["adm", "vif", "motion"]
-=======
-        features = ["adm", "vif", "motion", "float_ansnr"]
->>>>>>> 24bb5daf89 (docs: post-merge-train sweep — VMAFx + core/ path refs, ADR index, state.md)
         options = {
             "adm": {"debug": True},
             "vif": {"debug": True},

@@ -178,123 +178,6 @@ def _get_feature_key_for_score(self):
         return "psnr"
 
 
-class VmafLegacyQualityRunner(QualityRunner):
-    TYPE = "VMAF_legacy"
-
-    # VERSION = '1.1'
-    VERSION = "F" + VmafFeatureExtractor.VERSION + "-1.1"
-
-    FEATURE_ASSEMBLER_DICT = {"VMAF_feature": "all"}
-
-    FEATURE_RESCALE_DICT = {
-        "VMAF_feature_vif_scores": (0.0, 1.0),
-        "VMAF_feature_adm_scores": (0.4, 1.0),
-        "VMAF_feature_ansnr_scores": (10.0, 50.0),
-        "VMAF_feature_motion_scores": (0.0, 20.0),
-    }
-
-    SVM_MODEL_FILE = VmafConfig.model_path("other_models", "model_V8a.model")
-
-    # model_v8a.model is trained with customized feature order:
-    SVM_MODEL_ORDERED_SCORES_KEYS = [
-        "VMAF_feature_vif_scores",
-        "VMAF_feature_adm_scores",
-        "VMAF_feature_ansnr_scores",
-        "VMAF_feature_motion_scores",
-    ]
-
-    def _get_quality_scores(self, asset):
-        raise NotImplementedError
-
-    def _generate_result(self, asset):
-        raise NotImplementedError
-
-    def _get_vmaf_feature_assembler_instance(self, asset):
-        vmaf_fassembler = FeatureAssembler(
-            feature_dict=self.FEATURE_ASSEMBLER_DICT,
-            feature_option_dict=None,
-            assets=[asset],
-            logger=self.logger,
-            fifo_mode=self.fifo_mode,
-            delete_workdir=self.delete_workdir,
-            result_store=self.result_store,
-            optional_dict=None,
-            optional_dict2=None,
-            parallelize=False,  # parallelization already in a higher level
-            save_workfiles=self.save_workfiles,
-        )
-        return vmaf_fassembler
-
-    @override(Executor)
-    def _run_on_asset(self, asset):
-        # override Executor._run_on_asset(self, asset), which runs a
-        # FeatureAssembler, collect a feature vector, run
-        # TrainTestModel.predict() on it, and return a Result object
-        # (in this case, both Executor._run_on_asset(self, asset) and
-        # QualityRunner._read_result(self, asset) get bypassed.
-
-        vmaf_fassembler = self._get_vmaf_feature_assembler_instance(asset)
-        vmaf_fassembler.run()
-        feature_result = vmaf_fassembler.results[0]
-
-        # =====================================================================
-
-        # SVR predict
-        model = svmutil.svm_load_model(self.SVM_MODEL_FILE)
-
-        ordered_scaled_scores_list = []
-        for scores_key in self.SVM_MODEL_ORDERED_SCORES_KEYS:
-            scaled_scores = self._rescale(
-                feature_result[scores_key], self.FEATURE_RESCALE_DICT[scores_key]
-            )
-            ordered_scaled_scores_list.append(scaled_scores)
-
-        scores = []
-        for score_vector in zip(*ordered_scaled_scores_list):
-            vif, adm, ansnr, motion = score_vector
-            xs = [[vif, adm, ansnr, motion]]
-            score = svmutil.svm_predict([0], xs, model)[0][0]
-            score = self._post_correction(motion, score)
-            scores.append(score)
-
-        result_dict = {}
-        # add all feature result
-        result_dict.update(feature_result.result_dict)
-        # add quality score
-        result_dict[self.get_scores_key()] = scores
-
-        return Result(asset, self.executor_id, result_dict)
-
-    def _post_correction(self, motion, score):
-        # post-SVM correction
-        if motion > 12.0:
-            val = motion
-            if val > 20.0:
-                val = 20
-            score *= (val - 12) * 0.015 + 1
-        if score > 100.0:
-            score = 100.0
-        elif score < 0.0:
-            score = 0.0
-        return score
-
-    @classmethod
-    def _rescale(cls, vals, lower_upper_bound):
-        lower_bound, upper_bound = lower_upper_bound
-        vals = np.double(vals)
-        vals = np.clip(vals, lower_bound, upper_bound)
-        vals = (vals - lower_bound) / (upper_bound - lower_bound)
-        return vals
-
-    @override(Executor)
-    def _remove_result(self, asset):
-        # Override Executor._remove_result by redirecting it to the
-        # FeatureAssembler.
-
-        vmaf_fassembler = self._get_vmaf_feature_assembler_instance(asset)
-        vmaf_fassembler.remove_results()
-
-
 class VmafQualityRunnerModelMixin(object):
 
     def _load_model(self, asset):

@@ -0,0 +1,74 @@
+# ADR-0749: Sunset VmafLegacyQualityRunner (float-path runner)
+
+- **Status**: Accepted
+- **Date**: 2026-05-28
+- **Deciders**: lusoris
+- **Tags**: `python`, `quality-runner`, `breaking-change`, `cleanup`
+
+## Context
+
+`VmafLegacyQualityRunner` is a Python quality runner that drives the
+`VmafFeatureExtractor` float-path via the `vmafexec` binary, collecting
+the four legacy SVM features (vif, adm, ansnr, motion) and scoring them
+with a libsvm model (`model_V8a.model`).
+
+PR #38 removed the `float_ansnr` feature extractor from the C backend — it
+was a pre-VMAF metric that Netflix never adopted in production. After that
+removal, `VmafFeatureExtractor.ATOM_FEATURES_TO_VMAFEXEC_KEY_DICT` maps the
+`"ansnr"` feature to the key `"float_ansnr"`, which no longer appears in
+vmafexec output. Any invocation of `VmafLegacyQualityRunner` against the
+current binary silently returns no ansnr scores or raises a `KeyError` during
+the SVM scoring step.
+
+CI infrastructure agent T-LEGACY-RUNNER-ANSNR-BROKEN (PR #86) surfaced this
+as a required-status-check failure. The user authorized formal sunset of the
+runner class and all tests that exclusively exercise it.
+
+`VmafIntegerFeatureExtractor` (which inherits from `VmafFeatureExtractor` but
+uses the integer-path vmafexec keys) is unaffected; the canonical Netflix
+golden test `test_run_vmaf_runner` uses `VmafQualityRunner` and is untouched.
+
+## Decision
+
+Remove `VmafLegacyQualityRunner` from `compat/python-vmaf/core/quality_runner.py`
+and delete all Python tests that exclusively exercise it. `VmafFeatureExtractor`
+(the Python class) is retained because `VmafIntegerFeatureExtractor` and
+several other non-legacy quality runners depend on it.
+
+## Alternatives considered
+
+| Option | Pros | Cons | Why not chosen |
+|---|---|---|---|
+| Restore `float_ansnr` C implementation | Runner would work again | Reinstates a pre-VMAF metric Netflix never shipped; contradicts PR #38 rationale | Rejected — the feature was dropped deliberately |
+| Keep runner, mark deprecated, skip ansnr | Runner callable without crashing | Ansnr column silently absent; score math is wrong (SVM model expects 4 features); deceptive API | Rejected — broken API is worse than no API |
+| Sunset now (chosen) | Removes broken surface; unblocks CI | BREAKING change for any caller using the legacy Python runner | Accepted — callers should use `VmafQualityRunner` + a modern `.json` model |
+
+## Consequences
+
+- **Positive**: CI gate T-LEGACY-RUNNER-ANSNR-BROKEN resolved. Test suite
+  no longer has an unconditionally-failing test class that blocks the merge
+  train. The codebase no longer exports a broken public API.
+- **Negative**: Any Python code calling `VmafLegacyQualityRunner` will
+  receive an `ImportError` on the next install. Migration path: use
+  `VmafQualityRunner` with `vmaf_float_v0.6.1.pkl` or a current `.json` model.
+- **Neutral / follow-ups**:
+  - `VmafFeatureExtractor.ATOM_FEATURES` still lists `"ansnr"` (mapped to
+    the absent `float_ansnr` key). This is a latent hazard for any direct
+    caller. A follow-up PR should remove `ansnr` from `ATOM_FEATURES` per
+    the roadmap in `docs/research/0733-feature-importance-audit-2026-05-28.md`
+    Phase 2.
+  - `feature_extractor_test.py` contains many tests using `VmafFeatureExtractor`
+    directly, some asserting `ansnr` scores via the float path. Those tests
+    may also be broken; they are out of scope for this PR (they reference
+    features beyond just `ansnr` and require broader assessment).
+
+## References
+
+- PR #38: drop legacy `ansnr` feature extractor from C backend
+- PR #86: CI infrastructure agent flagged T-LEGACY-RUNNER-ANSNR-BROKEN
+- `docs/research/0733-feature-importance-audit-2026-05-28.md` — Phase 2
+  roadmap (remove `ansnr` from `VmafFeatureExtractor.ATOM_FEATURES`)
+- req: "Sunset the legacy `VmafFeatureExtractor` (float-path) runner.
+  PR #86's CI infra agent flagged this as T-LEGACY-RUNNER-ANSNR-BROKEN —
+  the legacy runner still calls `float_ansnr` which was dropped by PR #38.
+  User authorized formal sunset."
@@ -0,0 +1,94 @@
+# Research-0749: Sunset VmafLegacyQualityRunner
+
+**Date**: 2026-05-28
+**Author**: Agent (claude-sonnet-4-6)
+**Status**: Complete
+
+---
+
+## Summary
+
+`VmafLegacyQualityRunner` is a Python quality-runner class that drove the
+pre-2019 VMAF scoring pipeline via an SVM model (`model_V8a.model`) trained
+on four float features: vif, adm, ansnr, motion. After PR #38 dropped
+`float_ansnr` from the C feature extractor registry, any invocation of the
+runner results in a broken or empty result. CI agent T-LEGACY-RUNNER-ANSNR-BROKEN
+(surfaced in PR #86) confirmed the runner was unconditionally broken.
+
+---
+
+## What was removed
+
+| Artifact | Location | Lines removed |
+|---|---|---|
+| `VmafLegacyQualityRunner` class | `compat/python-vmaf/core/quality_runner.py` | ~115 |
+| `test_executor_id` (legacy runner version) | `python/test/quality_runner_test.py` | ~10 |
+| `test_run_vmaf_legacy_runner` | `python/test/quality_runner_test.py` | ~25 |
+| `test_run_vmaf_legacy_runner_10le` | `python/test/quality_runner_test.py` | ~25 |
+| `test_run_vmaf_legacy_runner_12le` | `python/test/quality_runner_test.py` | ~30 |
+| `test_run_vmaf_legacy_runner_with_result_store` | `python/test/quality_runner_test.py` | ~35 |
+| `ResultTest` class (legacy runner backed) | `python/test/result_test.py` | ~155 |
+| `ResultStoreTest` class (legacy runner backed) | `python/test/result_test.py` | ~40 |
+
+**Total**: approximately 435 lines of broken code removed.
+
+---
+
+## What was NOT removed
+
+- `VmafFeatureExtractor` Python class — retained. Used by
+  `VmafIntegerFeatureExtractor` (which maps to integer-path vmafexec keys),
+  `VmafQualityRunner`, `BaggingVmafQualityRunner`, and numerous tests in
+  `feature_extractor_test.py`.
+- All `VmafIntegerFeatureExtractor` tests — unaffected.
+- All canonical Netflix golden tests (`test_run_vmaf_runner` and checkerboard
+  tests) — protected and untouched per CLAUDE §1.
+- `ResultFormattingTest`, `ResultStoreTestWithNone`, `ResultAggregatingTest`,
+  `ScoreAggregationTest` in `result_test.py` — these use `VmafQualityRunner`
+  or static fixture files and are unaffected.
+
+---
+
+## Root cause
+
+PR #38 deleted `core/src/feature/float_ansnr.c` and removed
+`vmaf_fex_float_ansnr` from the C feature registry. `VmafFeatureExtractor`
+in Python maps `"ansnr"` to `"float_ansnr"` in its
+`ATOM_FEATURES_TO_VMAFEXEC_KEY_DICT`. When vmafexec no longer emits a
+`float_ansnr` attribute, the XML parse produces no entry for that key,
+and the SVM scoring step in `VmafLegacyQualityRunner._run_on_asset` either
+returns zero scores or raises a `KeyError` depending on call path.
+
+---
+
+## Before / after test counts
+
+| File | Before | After | Delta |
+|---|---|---|---|
+| `quality_runner_test.py` | 5 removed tests + rest | rest only | -5 tests |
+| `result_test.py` | `ResultTest` (2) + `ResultStoreTest` (1) + rest | rest only | -3 tests |
+
+Canonical Netflix golden test (`test_run_vmaf_runner`) status: PASS (unmodified).
+
+---
+
+## Follow-up work
+
+Per `docs/research/0733-feature-importance-audit-2026-05-28.md` Phase 2:
+
+1. Remove `"ansnr"` from `VmafFeatureExtractor.ATOM_FEATURES` and its
+   `ATOM_FEATURES_TO_VMAFEXEC_KEY_DICT` entry.
+2. Assess `feature_extractor_test.py` tests that use `VmafFeatureExtractor`
+   and assert `ansnr` scores — those tests may be broken against the current
+   binary.
+3. Remove `float_ansnr` from HIP backend scaffolding (`float_ansnr_hip`)
+   which still has references in `docs/rebase-notes.md`.
+
+---
+
+## References
+
+- PR #38: drop `float_ansnr` from C backend
+- PR #86: CI infra agent T-LEGACY-RUNNER-ANSNR-BROKEN
+- ADR-0749: formal sunset decision
+- `docs/research/0733-feature-importance-audit-2026-05-28.md`