feat(diffctx): adaptive scoring modes (auto/precise/discover)

Add --scoring flag with three modes:
- auto (default): adapts to repo size (PPR for small, EgoGraph for large)
- precise: PPR scoring, best precision on small repos
- discover: EgoGraph 2-hop BFS, 52% nontrivial recall on ContextBench

PipelineConfig centralizes all algorithm parameters per mode.
Auto mode uses fragment count threshold (300) to switch.
YAML tests: 83.7% avg, 1403 passed (auto → PPR for small test repos).
Benchmarks with discover: 52% CB nontrivial, 25% LOO recall.

Author: nikolay-e

src/treemapper/cli.py

@@ -311,6 +311,12 @@ def _build_main_parser() -> argparse.ArgumentParser:
         metavar="F",
         help="Minimum relevance to include a fragment (default: 0.08, lower = more context)",
     )
+    diff_group.add_argument(
+        "--scoring",
+        choices=["auto", "precise", "discover"],
+        default="auto",
+        help="Scoring mode: auto (adapts to repo size), precise (PPR, best for small repos), discover (ego-graph, best for cross-file)",
+    )
     diff_group.add_argument(
         "--full",
         action="store_true",

src/treemapper/diffctx/mode.py

@@ -0,0 +1,50 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+from enum import Enum
+
+
+class ScoringMode(Enum):
+    AUTO = "auto"
+    PRECISE = "precise"
+    DISCOVER = "discover"
+
+
+@dataclass(frozen=True)
+class PipelineConfig:
+    discovery: str
+    scoring: str
+    low_relevance: bool
+    bm25_top_k: int
+    ego_depth: int
+    ppr_alpha: float
+
+    @staticmethod
+    def from_mode(mode: ScoringMode, n_fragments: int = 0) -> PipelineConfig:
+        if mode == ScoringMode.PRECISE:
+            return PipelineConfig(
+                discovery="default",
+                scoring="ppr",
+                low_relevance=True,
+                bm25_top_k=0,
+                ego_depth=1,
+                ppr_alpha=0.60,
+            )
+        if mode == ScoringMode.DISCOVER:
+            return PipelineConfig(
+                discovery="ensemble",
+                scoring="ego",
+                low_relevance=False,
+                bm25_top_k=1,
+                ego_depth=2,
+                ppr_alpha=0.60,
+            )
+        is_large = n_fragments > 300
+        return PipelineConfig(
+            discovery="ensemble" if is_large else "default",
+            scoring="ego" if is_large else "ppr",
+            low_relevance=not is_large,
+            bm25_top_k=1 if is_large else 0,
+            ego_depth=2 if is_large else 1,
+            ppr_alpha=0.60,
+        )

src/treemapper/diffctx/pipeline.py

@@ -16,9 +16,19 @@
 from .file_importance import compute_file_importance
 from .fragmentation import _process_files_for_fragments
 from .git import CatFileBatch, GitError, split_diff_range
+from .mode import PipelineConfig, ScoringMode
 from .postpass import _coherence_post_pass, _ensure_changed_files_represented
 from .render import build_diff_context_output
-from .scoring import DiscoveryContext, EgoGraphScoring, EnsembleDiscovery, PPRScoring, ScoringStrategy
+from .scoring import (
+    BM25Discovery,
+    DefaultDiscovery,
+    DiscoveryContext,
+    DiscoveryStrategy,
+    EgoGraphScoring,
+    EnsembleDiscovery,
+    PPRScoring,
+    ScoringStrategy,
+)
 from .select import lazy_greedy_select
 from .signatures import _generate_signature_variants
 from .types import Fragment, FragmentId
@@ -151,6 +161,12 @@ def _log_ppr_mode(
     )
 
 
+def _create_discovery(config: PipelineConfig) -> DiscoveryStrategy:
+    if config.discovery == "ensemble":
+        return EnsembleDiscovery([DefaultDiscovery(), BM25Discovery(top_k=config.bm25_top_k)])
+    return DefaultDiscovery()
+
+
 def _empty_tree(root_dir: Path) -> dict[str, Any]:
     return {
         "name": root_dir.name,
@@ -171,6 +187,7 @@ def build_diff_context(
     no_default_ignores: bool = False,
     full: bool = False,
     whitelist_file: Path | None = None,
+    scoring_mode: str = "auto",
 ) -> dict[str, Any]:
     _validate_inputs(root_dir, alpha, tau, budget_tokens)
     root_dir = root_dir.resolve()
@@ -217,6 +234,9 @@ def build_diff_context(
             except (OSError, UnicodeDecodeError):
                 continue
 
+        mode = ScoringMode(os.environ.get("DIFFCTX_SCORING", scoring_mode))
+        config = PipelineConfig.from_mode(mode, n_fragments=len(all_fragments))
+
         discovery_ctx = DiscoveryContext(
             root_dir=root_dir,
             changed_files=changed_files,
@@ -225,8 +245,7 @@ def build_diff_context(
             expansion_concepts=frozenset(expansion_concepts),
             file_cache=file_cache,
         )
-        discovery_strategy = EnsembleDiscovery()
-        discovered_files = discovery_strategy.discover(discovery_ctx)
+        discovered_files = _create_discovery(config).discover(discovery_ctx)
         discovered_files = [_normalize_path(p, root_dir) for p in discovered_files]
         all_fragments.extend(
             _process_files_for_fragments(discovered_files, root_dir, preferred_revs, seen_frag_ids, batch_reader)
@@ -275,7 +294,9 @@ def build_diff_context(
             hunks=hunks,
             repo_root=root_dir,
             seed_weights=seed_weights,
-            scoring_strategy=EgoGraphScoring() if os.environ.get("DIFFCTX_SCORING") == "ego" else PPRScoring(alpha=alpha),
+            scoring_strategy=(
+                EgoGraphScoring(max_depth=config.ego_depth) if config.scoring == "ego" else PPRScoring(alpha=config.ppr_alpha)
+            ),
             discovered_paths=set(discovered_files),
         )
         effective_budget = budget_tokens if budget_tokens is not None else _UNLIMITED_BUDGET

src/treemapper/treemapper.py

@@ -32,6 +32,7 @@ def _build_diff_tree(args: ParsedArgs) -> dict[str, Any]:
             no_default_ignores=args.no_default_ignores,
             full=args.full_diff,
             whitelist_file=args.whitelist_file,
+            scoring_mode=getattr(args, "scoring", "auto"),
         )
     except GitError as e:
         logger.error("%s", e)

whitelist_vulture.py

@@ -27,6 +27,7 @@
 )
 from treemapper.diffctx.graph import Graph
 from treemapper.diffctx.graph_analytics import QuotientNode
+from treemapper.diffctx.mode import PipelineConfig, ScoringMode
 from treemapper.diffctx.project_graph import ProjectGraph
 from treemapper.diffctx.scoring import EgoGraphScoring
 from treemapper.diffctx.tokenizer import detect_profile, is_nlp_available
@@ -63,4 +64,6 @@
 SqlEdgeBuilder
 ZigEdgeBuilder
 blast_radius = graph_analytics.blast_radius
+ScoringMode.AUTO
+PipelineConfig.low_relevance
 EgoGraphScoring