fix: drop unreachable low_relevance_threshold gate and its false eval metadata

Author: nikolay-e

CHANGELOG.md

@@ -41,6 +41,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   exercise git's real rename path instead of silently degrading to adds
   (#176).
 
+### Removed
+
+- The dead `low_relevance_threshold` gate.
+  `PipelineConfig.low_relevance_filter` was `false` for every scoring mode, so
+  `filter_low_relevance` — and with it `FILTERING.low_relevance_threshold`
+  (0.015) and its `size_penalty_*` scaling — could never execute;
+  `filter_positive_relevance` was always the gate that ran. ContextBench result
+  rows nevertheless stamped `low_relevance_threshold: 0.015` into their `config`
+  block, so every recorded run attributed its selections to a threshold that
+  never fired. Removed on both sides. Bit-equivalent: identical output across
+  four scoring modes × three diff ranges, plus the full 2725-case corpus.
+
 ## [1.12.3] - 2026-07-27
 
 ### Added

crates/diffctx-native/src/config/filtering.rs

@@ -8,9 +8,6 @@ pub struct FilteringConfig {
     pub definition_proximity_half_decay: f64,
     pub hub_reverse_threshold: usize,
     pub max_context_fragments_per_file: usize,
-    pub low_relevance_threshold: f64,
-    pub size_penalty_base_tokens: f64,
-    pub size_penalty_exponent: f64,
 }
 
 impl Default for FilteringConfig {
@@ -21,9 +18,6 @@ impl Default for FilteringConfig {
             definition_proximity_half_decay: 5.0,
             hub_reverse_threshold: 2,
             max_context_fragments_per_file: 30,
-            low_relevance_threshold: 0.015,
-            size_penalty_base_tokens: 100.0,
-            size_penalty_exponent: 0.5,
         }
     }
 }

crates/diffctx-native/src/filtering.rs

@@ -35,13 +35,6 @@ fn proximity_score(frag: &Fragment, file_hunks: &[(u32, u32)]) -> f64 {
     FILTERING.proximity_floor_max / (1.0 + min_gap as f64 / half_decay)
 }
 
-fn effective_relevance_threshold(token_count: u32) -> f64 {
-    let size_factor = (token_count as f64 / FILTERING.size_penalty_base_tokens)
-        .max(1.0)
-        .powf(FILTERING.size_penalty_exponent);
-    FILTERING.low_relevance_threshold * size_factor
-}
-
 pub fn apply_hunk_proximity_bonus(
     rel: &mut FxHashMap<FragmentId, f64>,
     core_ids: &FxHashSet<FragmentId>,
@@ -243,21 +236,6 @@ pub fn filter_unrelated_fragments(
         .collect()
 }
 
-pub fn filter_low_relevance(
-    fragments: Vec<Fragment>,
-    core_ids: &FxHashSet<FragmentId>,
-    rel: &FxHashMap<FragmentId, f64>,
-) -> Vec<Fragment> {
-    fragments
-        .into_iter()
-        .filter(|f| {
-            core_ids.contains(&f.id)
-                || rel.get(&f.id).copied().unwrap_or(0.0)
-                    >= effective_relevance_threshold(f.token_count)
-        })
-        .collect()
-}
-
 pub fn filter_positive_relevance(
     fragments: Vec<Fragment>,
     core_ids: &FxHashSet<FragmentId>,

crates/diffctx-native/src/mode.rs

@@ -57,7 +57,6 @@ pub struct PipelineConfig {
     pub discovery: DiscoveryKind,
     pub scoring: ScoringKind,
     pub objective: ObjectiveMode,
-    pub low_relevance_filter: bool,
     pub bm25_top_k: usize,
     pub ego_depth: usize,
     pub ppr_alpha: f64,
@@ -70,7 +69,6 @@ impl PipelineConfig {
             ScoringMode::Ppr => Self {
                 discovery: DiscoveryKind::Ensemble,
                 scoring: ScoringKind::Ppr,
-                low_relevance_filter: false,
                 bm25_top_k: m.bm25_top_k_primary,
                 ego_depth: m.ego_depth_default,
                 ppr_alpha: PPR.alpha,
@@ -79,7 +77,6 @@ impl PipelineConfig {
             ScoringMode::Ego => Self {
                 discovery: DiscoveryKind::Ensemble,
                 scoring: ScoringKind::Ego,
-                low_relevance_filter: false,
                 bm25_top_k: m.bm25_top_k_primary,
                 ego_depth: m.ego_depth_extended,
                 ppr_alpha: PPR.alpha,
@@ -88,7 +85,6 @@ impl PipelineConfig {
             ScoringMode::Bm25 => Self {
                 discovery: DiscoveryKind::Ensemble,
                 scoring: ScoringKind::Bm25,
-                low_relevance_filter: false,
                 bm25_top_k: m.bm25_top_k_off,
                 ego_depth: m.ego_depth_default,
                 ppr_alpha: PPR.alpha,
@@ -100,7 +96,6 @@ impl PipelineConfig {
             ScoringMode::Rrf => Self {
                 discovery: DiscoveryKind::Ensemble,
                 scoring: ScoringKind::Rrf,
-                low_relevance_filter: false,
                 bm25_top_k: m.bm25_top_k_primary,
                 ego_depth: m.ego_depth_extended,
                 ppr_alpha: PPR.alpha,

crates/diffctx-native/src/scoring.rs

@@ -34,10 +34,7 @@ pub struct ScoringResult {
 pub fn create_scoring_strategy(config: &PipelineConfig) -> Box<dyn ScoringStrategy> {
     match config.scoring {
         ScoringKind::Ego => Box::new(EgoGraphScoring::new(config.ego_depth)),
-        ScoringKind::Ppr => Box::new(PPRScoring::new(
-            config.ppr_alpha,
-            config.low_relevance_filter,
-        )),
+        ScoringKind::Ppr => Box::new(PPRScoring::new(config.ppr_alpha)),
         ScoringKind::Bm25 => Box::new(BM25Scoring),
         ScoringKind::Rrf => Box::new(RrfFusionScoring::new(config.ego_depth)),
     }
@@ -57,15 +54,11 @@ pub trait ScoringStrategy: Send + Sync {
 
 pub struct PPRScoring {
     pub alpha: f64,
-    pub low_relevance_filter: bool,
 }
 
 impl PPRScoring {
-    pub fn new(alpha: f64, low_relevance_filter: bool) -> Self {
-        Self {
-            alpha,
-            low_relevance_filter,
-        }
+    pub fn new(alpha: f64) -> Self {
+        Self { alpha }
     }
 }
 
@@ -104,11 +97,7 @@ impl ScoringStrategy for PPRScoring {
         filtering::apply_hunk_proximity_bonus(&mut rel_scores, core_ids, all_fragments, hunks);
 
         let filtered = filtering::filter_unrelated_fragments(all_fragments, core_ids, &g);
-        let filtered = if self.low_relevance_filter {
-            filtering::filter_low_relevance(filtered, core_ids, &rel_scores)
-        } else {
-            filtering::filter_positive_relevance(filtered, core_ids, &rel_scores)
-        };
+        let filtered = filtering::filter_positive_relevance(filtered, core_ids, &rel_scores);
         let filtered = filtering::cap_context_fragments(filtered, core_ids, &rel_scores);
 
         ScoringResult {

eval/workflows/contextbench.py

@@ -170,7 +170,6 @@ def run_baseline_bm25(repo_dir: Path, budget: int = 8000) -> dict | None:
 
 def _get_diffctx_config() -> dict:
     return {
-        "low_relevance_threshold": 0.015,
         "proximity_decay": 0.30,
         "peripheral_cap": 0.15,
         "max_context_frags_per_file": 30,