python-delphi-lsp 3.2.0__tar.gz → 3.2.1__tar.gz

{python_delphi_lsp-3.2.0/python_delphi_lsp.egg-info → python_delphi_lsp-3.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: python-delphi-lsp
-Version: 3.2.0
+Version: 3.2.1
 Summary: Python Delphi/Object Pascal parser, semantic indexer, and language server.
 Author: Dark Light
 License-Expression: MPL-2.0
@@ -45,7 +45,7 @@ Dynamic: license-file
 
 `python-delphi-lsp` parses Delphi/Object Pascal, builds semantic and project
 indexes, serves LSP, and provides bounded codebase navigation for agents.
-Version 3.2.0 is authored by Dark Light and supports Windows, macOS, and Linux.
+Version 3.2.1 is authored by Dark Light and supports Windows, macOS, and Linux.
 
 ## Install and quick start
 
@@ -148,7 +148,10 @@ The root `opencode.json` starts the installed package portably:
     "delphi": {
       "command": ["delphi-lsp"],
       "extensions": [".pas", ".dpr", ".dpk", ".inc"],
-      "initialization": {"autoDiscoverPaths": true}
+      "initialization": {
+        "autoDiscoverPaths": true,
+        "semanticCacheSize": 512
+      }
     }
   }
 }
@@ -158,6 +161,10 @@ The root `opencode.json` starts the installed package portably:
 environment section; LSP remains available for normal editor and OpenCode use,
 including large sources.
 
+`semanticCacheSize` limits retained per-file outline models while keeping
+source text available. Its default is 512; set it to zero to disable retained
+snapshot models. The value is preserved when workspace folders change.
+
 The LSP builds its structural index through the same optimized outline path for
 every source, with no file-size threshold. Definition, hover, references,
 rename, completion, document symbols, workspace symbols, and diagnostics remain
@@ -203,7 +210,7 @@ delphi-lsp-agent view --root PATH [--project-file FILE] --layer LAYER
                       [--query TEXT] [--format markdown|json] [--deep-projects]
                       [--workers auto|N]
 delphi-lsp-agent index --root PATH [--project-file FILE] [--out FILE]
-                       [--workers auto|N]
+                       [--workers auto|N] [--deep-projects]
 delphi-lsp-agent wiki export --root PATH [--project-file FILE] [--out DIRECTORY]
                              [--workers auto|N] [--force] [--quiet]
 delphi-lsp-agent query --root PATH ACTION [VALUE]
@@ -218,6 +225,13 @@ delphi-lsp-agent opencode install [--target PATH] [--python PYTHON]
 delphi-lsp-agent worker --root PATH [--project-file FILE] [--workers auto|N]
 ```
 
+`--deep-projects` opts into per-project dependency parsing for the `projects`
+layer. It is off by default because it dominates run time on real repositories:
+on a 60-project corpus it cost 36.9 s against 1.0 s for the same index, and it
+adds nothing to the unit model set. Without it the `projects` and `problems`
+layers report `deep_indexed: false`, and `index` records `deep_projects` at the
+top level of its JSON, so a shallow index is never mistaken for a complete one.
+
 The `cache` commands manage one daemon per canonical root. Use these:
 
 ```bash

{python_delphi_lsp-3.2.0 → python_delphi_lsp-3.2.1}/README.md

@@ -2,7 +2,7 @@
 
 `python-delphi-lsp` parses Delphi/Object Pascal, builds semantic and project
 indexes, serves LSP, and provides bounded codebase navigation for agents.
-Version 3.2.0 is authored by Dark Light and supports Windows, macOS, and Linux.
+Version 3.2.1 is authored by Dark Light and supports Windows, macOS, and Linux.
 
 ## Install and quick start
 
@@ -105,7 +105,10 @@ The root `opencode.json` starts the installed package portably:
     "delphi": {
       "command": ["delphi-lsp"],
       "extensions": [".pas", ".dpr", ".dpk", ".inc"],
-      "initialization": {"autoDiscoverPaths": true}
+      "initialization": {
+        "autoDiscoverPaths": true,
+        "semanticCacheSize": 512
+      }
     }
   }
 }
@@ -115,6 +118,10 @@ The root `opencode.json` starts the installed package portably:
 environment section; LSP remains available for normal editor and OpenCode use,
 including large sources.
 
+`semanticCacheSize` limits retained per-file outline models while keeping
+source text available. Its default is 512; set it to zero to disable retained
+snapshot models. The value is preserved when workspace folders change.
+
 The LSP builds its structural index through the same optimized outline path for
 every source, with no file-size threshold. Definition, hover, references,
 rename, completion, document symbols, workspace symbols, and diagnostics remain
@@ -160,7 +167,7 @@ delphi-lsp-agent view --root PATH [--project-file FILE] --layer LAYER
                       [--query TEXT] [--format markdown|json] [--deep-projects]
                       [--workers auto|N]
 delphi-lsp-agent index --root PATH [--project-file FILE] [--out FILE]
-                       [--workers auto|N]
+                       [--workers auto|N] [--deep-projects]
 delphi-lsp-agent wiki export --root PATH [--project-file FILE] [--out DIRECTORY]
                              [--workers auto|N] [--force] [--quiet]
 delphi-lsp-agent query --root PATH ACTION [VALUE]
@@ -175,6 +182,13 @@ delphi-lsp-agent opencode install [--target PATH] [--python PYTHON]
 delphi-lsp-agent worker --root PATH [--project-file FILE] [--workers auto|N]
 ```
 
+`--deep-projects` opts into per-project dependency parsing for the `projects`
+layer. It is off by default because it dominates run time on real repositories:
+on a 60-project corpus it cost 36.9 s against 1.0 s for the same index, and it
+adds nothing to the unit model set. Without it the `projects` and `problems`
+layers report `deep_indexed: false`, and `index` records `deep_projects` at the
+top level of its JSON, so a shallow index is never mistaken for a complete one.
+
 The `cache` commands manage one daemon per canonical root. Use these:
 
 ```bash

python_delphi_lsp-3.2.1/delphi_lsp/_version.py

@@ -0,0 +1 @@
+__version__ = "3.2.1"

{python_delphi_lsp-3.2.0 → python_delphi_lsp-3.2.1}/delphi_lsp/agent_cli.py

@@ -93,6 +93,11 @@ def build_parser() -> argparse.ArgumentParser:
     index.add_argument("--project-file", type=Path)
     index.add_argument("--out", type=Path, default=Path(".delphi-lsp") / "agent-index" / "index.json")
     index.add_argument("--workers", type=parse_worker_setting, default=0)
+    index.add_argument(
+        "--deep-projects",
+        action="store_true",
+        help="Deep-parse project dependencies for the projects layer.",
+    )
     index.set_defaults(func=_index)
 
     wiki = subcommands.add_parser("wiki", help="Export a portable Markdown knowledge wiki.")
@@ -328,10 +333,11 @@ def _index(args: argparse.Namespace) -> None:
     index = build_codebase_index(
         args.root,
         project_file=args.project_file,
-        index_projects=True,
+        index_projects=args.deep_projects,
         workers=args.workers,
     )
     payload = {
+        "deep_projects": bool(args.deep_projects),
         "overview": layer_payload(index, "overview"),
         "projects": layer_payload(index, "projects"),
         "problems": layer_payload(index, "problems"),

{python_delphi_lsp-3.2.0 → python_delphi_lsp-3.2.1}/delphi_lsp/agent_layers.py

@@ -654,7 +654,12 @@ def _problems_payload(index: CodebaseIndex) -> dict[str, Any]:
                     "file": problem.file_name,
                 }
             )
-    return {"layer": "problems", "root": index.root, "items": items}
+    return {
+        "layer": "problems",
+        "root": index.root,
+        "items": items,
+        "deep_indexed": bool(index.project_results),
+    }
 
 
 def _metrics_payload(index: CodebaseIndex, *, query: str) -> dict[str, Any]:

{python_delphi_lsp-3.2.0 → python_delphi_lsp-3.2.1}/delphi_lsp/lsp_server.py

@@ -39,6 +39,7 @@ from .source_reader import read_source_text
 from .workspace import WorkspaceSemanticResult
 
 _WORKSPACE_SYMBOL_QUERY_CACHE_SIZE = 8
+_DEFAULT_SEMANTIC_CACHE_SIZE = 512
 _MAX_WORKSPACE_SYMBOLS = 1_000
 _DELPHI_IDENTIFIER_RE = re.compile(r"[A-Za-z_][A-Za-z0-9_]*\Z")
 _DELPHI_KEYWORDS = frozenset(keyword.casefold() for keyword in KEYWORDS)
@@ -70,6 +71,7 @@ class WorkspaceConfig:
     extensions: tuple[str, ...] = ('.pas', '.dpr', '.dpk', '.inc')
     eager_index: bool = False
     auto_discover_paths: bool = True
+    semantic_cache_size: int = _DEFAULT_SEMANTIC_CACHE_SIZE
     discovered_include_paths: list[str] = field(default_factory=list)
     discovered_search_paths: list[str] = field(default_factory=list)
     discovered_defines: list[str] = field(default_factory=list)
@@ -86,6 +88,10 @@ class LspWorkspaceState:
     )
     workspace_files: set[str] = field(default_factory=set)
     file_cache: dict[str, FileSnapshot] = field(default_factory=dict)
+    semantic_cache_lru: OrderedDict[str, None] = field(
+        default_factory=OrderedDict,
+    )
+    semantic_evicted_paths: set[str] = field(default_factory=set)
     config_warnings: list[str] = field(default_factory=list)
     project_configs: tuple[ProjectPathConfig, ...] = field(
         default_factory=tuple,
@@ -111,6 +117,8 @@ class LspWorkspaceState:
         self.config = configured
         self.workspace_files = set()
         self.file_cache = {}
+        self.semantic_cache_lru.clear()
+        self.semantic_evicted_paths.clear()
         self.workspace = None
         self.workspace_symbol_index = None
         self.workspace_symbol_query_cache.clear()
@@ -156,6 +164,7 @@ class LspWorkspaceState:
             extensions=config.extensions,
             eager_index=config.eager_index,
             auto_discover_paths=config.auto_discover_paths,
+            semantic_cache_size=config.semantic_cache_size,
             discovered_include_paths=discovered_include_paths,
             discovered_search_paths=discovered_search_paths,
             discovered_defines=discovered_defines,
@@ -270,6 +279,8 @@ class LspWorkspaceState:
         to_remove = [path for path in self.file_cache if path not in self.workspace_files]
         for path in to_remove:
             self.file_cache.pop(path, None)
+            self.semantic_cache_lru.pop(path, None)
+            self.semantic_evicted_paths.discard(path)
         for path in self.workspace_files:
             try:
                 mtime = os.path.getmtime(path)
@@ -282,6 +293,8 @@ class LspWorkspaceState:
                 text = read_source_text(Path(path))
             except (OSError, UnicodeError):
                 continue
+            self.semantic_cache_lru.pop(path, None)
+            self.semantic_evicted_paths.discard(path)
             self.file_cache[path] = FileSnapshot(path=path, text=text, mtime=mtime)
 
     def _collect_sources(self) -> dict[str, str]:
@@ -324,11 +337,15 @@ class LspWorkspaceState:
             text = read_source_text(path)
         except (OSError, UnicodeError):
             return None
-        return build_outline_semantic_model(
+        model = build_outline_semantic_model(
             text,
             file_name,
             defines=self.config.defines,
         )
+        cached = self.file_cache.get(file_name)
+        if cached is not None and cached.text == text:
+            self._retain_cached_semantic(file_name, cached, model)
+        return model
 
     def full_semantic_for_uri(self, uri: str) -> SemanticModel | None:
         file_name = uri_to_path(uri)
@@ -435,15 +452,19 @@ class LspWorkspaceState:
         if not sources:
             return None
         models: dict[str, SemanticModel] = {}
+        previous_models = self.workspace.models if self.workspace is not None else {}
+        pending: list[tuple[str, str, FileSnapshot | None]] = []
         for file_name, text in sources.items():
             cached = self.file_cache.get(file_name)
-            if (
-                cached is not None
-                and cached.text == text
-                and cached.semantic is not None
-            ):
-                models[file_name] = cached.semantic
-                continue
+            if cached is not None and cached.text == text:
+                model = cached.semantic
+                if model is None and file_name in self.semantic_evicted_paths:
+                    model = previous_models.get(file_name)
+                if model is not None:
+                    models[file_name] = model
+                    continue
+            pending.append((file_name, text, cached))
+        for file_name, text, cached in pending:
             model = build_outline_semantic_model(
                 text,
                 file_name,
@@ -451,7 +472,7 @@ class LspWorkspaceState:
             )
             models[file_name] = model
             if cached is not None and cached.text == text:
-                cached.semantic = model
+                self._retain_cached_semantic(file_name, cached, model)
         index = SymbolIndex()
         for model in models.values():
             index.register_unit(model.unit_scope.name, model.unit_scope)
@@ -459,6 +480,27 @@ class LspWorkspaceState:
             model.index = index
         return WorkspaceSemanticResult(models=models, index=index)
 
+    def _retain_cached_semantic(
+        self,
+        file_name: str,
+        snapshot: FileSnapshot,
+        model: SemanticModel,
+    ) -> None:
+        snapshot.semantic = model
+        self.semantic_evicted_paths.discard(file_name)
+        self._touch_cached_semantic(file_name)
+
+    def _touch_cached_semantic(self, file_name: str) -> None:
+        self.semantic_cache_lru[file_name] = None
+        self.semantic_cache_lru.move_to_end(file_name)
+        limit = max(0, self.config.semantic_cache_size)
+        while len(self.semantic_cache_lru) > limit:
+            evicted, _ = self.semantic_cache_lru.popitem(last=False)
+            snapshot = self.file_cache.get(evicted)
+            if snapshot is not None:
+                snapshot.semantic = None
+                self.semantic_evicted_paths.add(evicted)
+
     def model_for_path(self, path: str) -> SemanticModel | None:
         if self.workspace is None:
             return None
@@ -2247,12 +2289,21 @@ def create_server():
         search_paths = [uri_to_path(path) for path in init_opts.get('searchPaths', [])]
         defines = init_opts.get('defines', [])
         auto_discover_paths = init_opts.get('autoDiscoverPaths', True)
+        raw_semantic_cache_size = init_opts.get(
+            'semanticCacheSize',
+            _DEFAULT_SEMANTIC_CACHE_SIZE,
+        )
+        try:
+            semantic_cache_size = max(0, int(raw_semantic_cache_size))
+        except (TypeError, ValueError):
+            semantic_cache_size = _DEFAULT_SEMANTIC_CACHE_SIZE
         config = WorkspaceConfig(
             roots=roots,
             include_paths=include_paths,
             search_paths=search_paths,
             defines=defines,
             auto_discover_paths=bool(auto_discover_paths),
+            semantic_cache_size=semantic_cache_size,
         )
         state.configure(config)
         for warning in state.config_warnings:
@@ -2396,6 +2447,7 @@ def create_server():
                 extensions=state.config.extensions,
                 eager_index=state.config.eager_index,
                 auto_discover_paths=state.config.auto_discover_paths,
+                semantic_cache_size=state.config.semantic_cache_size,
             )
         )
         source_line_cache.clear()

{python_delphi_lsp-3.2.0 → python_delphi_lsp-3.2.1}/delphi_lsp/project_discovery.py

@@ -34,6 +34,9 @@ SKIP_DIRS = {
     ".worktrees",
     "node_modules",
 }
+_RAD_STUDIO_ALWAYS_SKIP_DIRS = {"__history", "__recovery"}
+_RAD_STUDIO_PROJECT_OUTPUT_DIRS = {"bin", "obj", "win32", "win64"}
+_DELPHI_PROJECT_MARKER_EXTENSIONS = {".dpr", ".dproj", ".dpk"}
 
 
 @dataclass(frozen=True)
@@ -102,8 +105,8 @@ def discover_delphi_project(
     )
     _emit_progress(on_progress, "discovery", str(root_path), 0, 0, None, "project discovery started")
 
-    seen_search: set[str] = set()
-    seen_include: set[str] = set()
+    seen_search: dict[str, str] = {}
+    seen_include: dict[str, str] = {}
     seen_defines: set[str] = set()
     seen_projects: set[str] = set()
     seen_configs: set[str] = set()
@@ -115,7 +118,7 @@ def discover_delphi_project(
 
     def add_path(
         target: list[str],
-        seen: set[str],
+        canonical_by_key: dict[str, str],
         origins: dict[str, list[str]],
         path: Path | str,
         *,
@@ -124,7 +127,7 @@ def discover_delphi_project(
     ) -> None:
         _add_resolved_path(
             target,
-            seen,
+            canonical_by_key,
             origins,
             str(path),
             base=base,
@@ -277,8 +280,8 @@ def populate_workspace_sources(
 ) -> DelphiProjectDiscovery:
     root_path = Path(discovery.root).expanduser().resolve()
     seen_sources = {source.casefold() for source in discovery.source_files}
-    seen_search = {path.casefold() for path in discovery.search_paths}
-    seen_include = {path.casefold() for path in discovery.include_paths}
+    seen_search = {path.casefold(): path for path in discovery.search_paths}
+    seen_include = {path.casefold(): path for path in discovery.include_paths}
 
     _scan_sources(root_path, discovery, seen_sources, on_progress=on_progress)
     total_sources = len(discovery.source_files)
@@ -526,7 +529,7 @@ def _read_dpr_paths(
     project: Path,
     discovery: DelphiProjectDiscovery,
     search_paths: list[str],
-    seen_search: set[str],
+    seen_search: dict[str, str],
     search_path_origins: dict[str, list[str]],
 ) -> None:
     try:
@@ -558,10 +561,10 @@ def _read_dproj(
     path: Path,
     discovery: DelphiProjectDiscovery,
     search_paths: list[str],
-    seen_search: set[str],
+    seen_search: dict[str, str],
     search_path_origins: dict[str, list[str]],
     include_paths: list[str],
-    seen_include: set[str],
+    seen_include: dict[str, str],
     include_path_origins: dict[str, list[str]],
     add_define,
 ) -> None:
@@ -619,10 +622,10 @@ def _read_cfg(
     path: Path,
     discovery: DelphiProjectDiscovery,
     search_paths: list[str],
-    seen_search: set[str],
+    seen_search: dict[str, str],
     search_path_origins: dict[str, list[str]],
     include_paths: list[str],
-    seen_include: set[str],
+    seen_include: dict[str, str],
     include_path_origins: dict[str, list[str]],
     add_define,
 ) -> None:
@@ -801,7 +804,7 @@ def _walk_files(
         directory_names[:] = sorted(
             name
             for name in directory_names
-            if name not in SKIP_DIRS
+            if not _skip_walk_directory(name, file_names)
             and (
                 project_config is None
                 or not project_config.excludes_workspace_path(current / name)
@@ -823,13 +826,27 @@ def _walk_files(
                 raise
 
 
+def _skip_walk_directory(name: str, sibling_file_names: list[str]) -> bool:
+    if name in SKIP_DIRS:
+        return True
+    normalized = name.casefold()
+    if normalized in _RAD_STUDIO_ALWAYS_SKIP_DIRS:
+        return True
+    if normalized not in _RAD_STUDIO_PROJECT_OUTPUT_DIRS:
+        return False
+    return any(
+        Path(file_name).suffix.casefold() in _DELPHI_PROJECT_MARKER_EXTENSIONS
+        for file_name in sibling_file_names
+    )
+
+
 def _is_path_too_long(error: OSError) -> bool:
     return getattr(error, "winerror", None) == 206 or error.errno == errno.ENAMETOOLONG
 
 
 def _add_resolved_path(
     target: list[str],
-    seen: set[str],
+    canonical_by_key: dict[str, str],
     origins: dict[str, list[str]],
     value: str,
     *,
@@ -846,10 +863,11 @@ def _add_resolved_path(
     ):
         return
     key = str(resolved).casefold()
-    if key not in seen:
-        seen.add(key)
-        target.append(str(resolved))
-    exposed_path = next(item for item in target if item.casefold() == key)
+    exposed_path = canonical_by_key.get(key)
+    if exposed_path is None:
+        exposed_path = str(resolved)
+        canonical_by_key[key] = exposed_path
+        target.append(exposed_path)
     _record_origin(origins, exposed_path, origin)
 
 

{python_delphi_lsp-3.2.0 → python_delphi_lsp-3.2.1}/docs/performance-delphi.md

@@ -125,16 +125,107 @@ is therefore:
 | Delphi realism | 2,983,643 | 5,053 | 497 | 2,366 | 3.230 s | 16.032 s |
 
 Clean discovery cost about 0.452 ms per source; realism cost about
-0.639 ms per source. The cold scaling exponents were 0.819 clean and 0.777
-with realism, both classified as linear by the harness. The earlier
-30.18-CPU-second pathological discovery result was therefore **not
-reproduced**. The overlay still made the known risk measurable: search paths
-grew 3.4×, sources 1.5×, and absolute discovery time 2.1×.
+0.639 ms per source. The earlier 30.18-CPU-second pathological discovery result
+was therefore **not reproduced**. The overlay still made the known risk
+measurable: search paths grew 3.4×, sources 1.5×, and absolute discovery time
+2.1×.
+
+The originally published cold scaling exponents — 0.819 clean and 0.777 with
+realism, both classified linear — were artefacts of fitting `time = b*n**k`
+without a constant term. Fixed per-run cost (interpreter start, imports, worker
+spawn) inflates the smallest slices and flattens the slope, which is how a
+workload that must read every line ended up with an exponent below 1: physically
+impossible. `fit_scaling` now estimates `time = a + b*n**k`. Refitted from the
+slice table above:
+
+| Profile | published k | refitted k | fixed cost a | r² | classification |
+|---|---:|---:|---:|---:|---|
+| Delphi clean | 0.819 | **1.1615** | 0.592 s | 0.99895 | n_log_n |
+| Delphi realism | 0.777 | 0.7640 | −0.244 s | 0.99598 | (not meaningful) |
+
+Clean is now a sensible reading: slightly worse than linear, with 0.592 s of
+fixed cost separated out. Realism still fits to a negative constant and a
+sub-linear exponent, which means the power law does not describe it — its cold
+curve is concave because the overlay's artifact cost is front-loaded relative to
+slice size rather than growing with it. Read the clean profile for scaling
+behaviour; read realism for the absolute cost of a realistic working tree.
 
 The parser audit intentionally covers the pinned base sources, not generated
 copies. It accepted all 3,416 sources; the role-aware clean ratios were 54.126%
 for units/projects and 95.916% for include fragments.
 
+## Shape-aware RAD pruning, 2026-07-29
+
+Discovery now skips `__history` and `__recovery` case-insensitively at every
+level. It skips `bin`, `obj`, `Win32`, and `Win64` only when the directory
+being walked contains a `.dpr`, `.dproj`, or `.dpk` file. The latter guard is
+important: FPC has genuine sources such as
+`packages/ptc/src/win32/base/*.inc`. `Debug`, `Release`, and `.vs` remain
+discoverable because the realism corpus measured no benefit from excluding
+them and those names have a meaningful false-positive risk.
+
+The table compares full fresh-process reports made from the same pinned
+corpora with the old and new directory predicate. Times are single report
+runs; the separate interleaved experiment below is the significance check.
+
+| Profile | Predicate | Visited | Indexed | Skipped | Search paths | Discovery | Cold | Warm | Peak RSS |
+|---|---|---:|---:|---:|---:|---:|---:|---:|---:|
+| FPC | old | 7,364 | 7,364 | 0 | 332 | 0.630 s | 11.073 s | 11.193 s | 508.5 MiB |
+| FPC | shape-aware | 7,364 | 7,364 | 0 | 332 | 0.692 s | 10.470 s | 10.755 s | 508.1 MiB |
+| Delphi clean | old | 3,416 | 3,416 | 0 | 692 | 1.430 s | 10.262 s | 10.663 s | 443.4 MiB |
+| Delphi clean | shape-aware | 3,416 | 3,416 | 0 | 692 | 1.419 s | 10.035 s | 10.382 s | 444.9 MiB |
+| Delphi realism | old | 5,053 | 5,053 | 0 | 2,366 | 3.332 s | 15.376 s | 15.141 s | 711.3 MiB |
+| Delphi realism | shape-aware | 5,053 | 3,672 | 1,381 | 1,945 | 2.832 s | 11.620 s | 11.561 s | 495.5 MiB |
+
+The realism overlay places its synthetic build copies at repository roots,
+which do not themselves contain project files. The literal shape rule
+therefore pruned the 1,176 history and 205 recovery sources, while retaining
+those synthetic `bin`/`obj`/`Win32`/`Win64` copies. A regression fixture
+separately proves that those four output names are pruned when they are
+siblings of a Delphi project file. This preserves the specified safety guard
+instead of weakening it to fit the overlay.
+
+Five alternating old/new discovery rounds, plus five alternating old/old null
+rounds, produced these medians:
+
+| Profile | Old | Shape-aware | Delta | Null noise | Discovery payload |
+|---|---:|---:|---:|---:|---|
+| FPC | 0.493 s | 0.497 s | +0.9% | 0.05% | byte-identical |
+| Delphi clean | 1.193 s | 1.188 s | -0.4% | 0.39% | byte-identical |
+| Delphi realism | 2.542 s | 2.277 s | -10.5% | 0.68% | intentionally 1,381 fewer sources |
+
+The FPC and clean timing differences are not treated as results. The realism
+improvement is over fifteen times its measured null noise. Its fresh-process
+cold result improved by 24.4%; the report's conservative 4x de-rating projects
+46.479 seconds on the target Windows i5, leaving 13.521 seconds (22.5%) of the
+60-second budget.
+
+All three parser audits were byte-identical before/after apart from their
+elapsed-time field. Accepted, unit-clean, and include-clean ratios were
+respectively 1.0/0.848995/0.744704 for FPC and
+1.0/0.682184/0.572650 for both Delphi reports.
+
+## Indexed path-origin lookup, 2026-07-29
+
+Resolved search and include paths now keep a `casefold` key to canonical
+appended-string index. The observable lists remain lists in insertion order;
+the dictionary only replaces the unconditional linear scan previously used to
+find the canonical spelling for origin recording.
+
+Five alternating linear/indexed rounds and five linear/linear null rounds
+produced:
+
+| Profile | Linear scan | Indexed lookup | Delta | Null noise | Search paths | Discovery payload |
+|---|---:|---:|---:|---:|---:|---|
+| FPC | 0.489 s | 0.436 s | -10.8% | 0.55% | 332 | byte-identical |
+| Delphi clean | 1.183 s | 1.046 s | -11.5% | 0.10% | 692 | byte-identical |
+| Delphi realism | 2.279 s | 1.187 s | -47.9% | 0.03% | 1,945 | byte-identical |
+
+The measured median reductions were 0.053 seconds for FPC, 0.137 seconds for
+Delphi clean, and 1.092 seconds for Delphi realism. Every complete discovery
+payload hash matched between implementations, including ordered
+`search_paths`, `include_paths`, and all origin maps.
+
 ## CI and next optimizations
 
 Fast tests use temporary fixtures only. The real smallest-slice probe is
@@ -146,16 +237,13 @@ DELPHI_LSP_DELPHI_CORPUS=/tmp/delphi-corpus \
   tests/test_delphi_performance_corpus.py
 ```
 
-The benchmark intentionally leaves these measured product changes for the
-next release:
+The benchmark intentionally leaves this semantic product change for a future
+release:
 
 1. derive overview `unit_count` directly from discovery/model inventory
    (`delphi_lsp/agent_layers.py:327`);
-2. replace the linear `next(...casefold())` lookup after every resolved path
-   with a keyed lookup (`delphi_lsp/project_discovery.py:849-854`);
-3. decide and implement Delphi-specific exclusions for `__history`,
-   `__recovery`, `bin`, `obj`, `Win32`, `Win64`, `Debug`, `Release`, and `.vs`
-   (`delphi_lsp/project_discovery.py:22-37`).
-
-Those changes are deliberately absent from these measurements so a follow-up
-can compare against the clean and realism reports above.
+
+The shape-aware exclusions and indexed path-origin lookup have now been
+implemented and measured above. The remaining item is deliberately absent
+from these measurements so follow-up work can compare against this
+checkpoint.

{python_delphi_lsp-3.2.0 → python_delphi_lsp-3.2.1}/docs/release-3.2.0.md

@@ -67,8 +67,10 @@ minor version 3.2.0 rather than a major-version bump.
 ## Verification
 
 The release is gated by two full-suite runs, repository-local Ruff and Mypy
-baselines, four real-corpus wiki exports, wheel/sdist validation, clean-room
-installation, and TestPyPI installation before production publication.
+baselines, four real-corpus wiki exports, wheel/sdist validation, and
+clean-room installation before and after production publication. TestPyPI was
+skipped under the release contract because no dedicated TestPyPI
+configuration was present.
 
 Install or upgrade with: