v0.2.3

Changed

• One route registration mints one Route node. @app.post('/x') used to
  satisfy both the flask method-shortcut detector and the fastapi holder
  detector, and framework is part of the Route id — so every such registration
  produced twin Route nodes. Framework is now claimed by per-file import
  evidence (files importing fastapi give it the app holder; .route
  stays flask; no-evidence @app.<verb> defaults to flask). On a real FastAPI
  repo every prior graph carried doubled routes.
• codingest build fails loudly on an empty result. A build whose graph
  contains no File and no Doc node now exits non-zero and writes NO
  artifact — previously it exited 0 and wrote a .kgl, the silence that once
  masked a walk-root bug. And a build rooted at a dot-prefixed directory
  (.hidden/) no longer derives module ids with an empty leading segment.
• Java line comments and Go interface methods are extracted again. Two
  match arms had gone silently dead under grammar renames (comment →
  line_comment/block_comment in tree-sitter-java 0.23; method_spec →
  method_elem in tree-sitter-go 0.25) — found by the new grammar-drift
  guard's first run. Java doc-comment gating and Go interface-method
  extraction resume on every Java/Go repo.
• Two new languages: Julia and R (seventeen total). Julia: functions in
  every definition form incl. multiple dispatch, structs with EXTENDS, nested
  modules, include("path.jl") file edges, using/import deliberately
  edge-free (namespace-shaped). R: both assignment shapes plus lambdas,
  conservative S4, source("path.R") file edges with .R/.r both
  registered, library()/require() deliberately edge-free. Each ships with
  a corpus, an additive golden, grammar-guard coverage, and collision-bait
  tests proving the namespace mechanisms abstain.
• Real-repo acceptance hardening across all languages. A five-agent sweep
  over real repositories (then a four-agent sweep for the long tail) produced
  three same-day fixes: Rust bare use paths resolve under the crate root and
  never reach the language-blind fallback (killing a false edge into a
  same-named Python package stub, and resolving pub use re-exports);
  multi-line route decorators — the dominant FastAPI style — no longer
  silently produce zero Route nodes; and the generic import fallback now
  claims File targets only for languages whose module coordinates name exactly
  one file (python/agc/dart) — namespace languages abstain, eliminating
  measured wrong-edge families (Java 438/438 false, PHP 146 src→tests
  inversions, C# 118 arbitrary picks).
• File→File import edges exist for Rust, Python, C/C++ and Dart. A
  five-defect family (independently reported by mcp-servers, 2026-08-14,
  upstream lead natashahirt/kglite PR #4) left the file-level dependency graph
  near-empty outside JS/TS: import strings were written in coordinate systems
  the resolver never converted from. Now: Rust crate::/super::/self::
  paths rewrite into the module coordinates files are stamped with (a 41-file
  scan that produced 0 File→File edges and 49 edges to one Module "crate"
  now resolves per-file); use foo as bar and use a::{b, c} are extracted;
  Python relative imports resolve against the importing file's package,
  import a, b/aliased/single-name-from origins all survive, and imports
  under if TYPE_CHECKING:/try:/function bodies are seen; Dart package:
  URIs keep their directory structure (a/x.dart and b/x.dart no longer
  collide) and relative URIs resolve file-relative; C/C++ angle includes are
  structurally excluded from file edges (an angle include naming a real
  project file no longer manufactures one — quoted-include resolution was
  already correct since 0.2.2). CALLS edges over these routes gain
  import_backed=true. Six goldens moved deliberately (four new import
  corpora + regenerations), each with its reason recorded per commit; a
  Python-ast oracle (machinery sharing nothing with the parsers) verifies
  the Python edge set at 7/7 recall.
• Engine floor moves to kglite 0.16.1. Two things reach us. (1) The wire
  break: kglite_value_to_json now emits real JSON for nine Value variants
  that previously leaked Rust Debug syntax — codingest query --format json
  returning a node/relationship/path/temporal now prints a structured object
  ({"id":…,"labels":[…],"properties":{…}}) instead of a quoted Debug string.
  Verified live against a built graph; anything parsing the old strings must
  move. (2) The docs-pass column_value copy only converts List/Map of
  scalars, and every parity corpus proves the nine variants unreachable there —
  goldens are byte-identical across the bump. Also in 0.16.1: all four
  NodeData-sentinel documentation asks codingest filed on 2026-08-14 shipped
  (rustdoc corrected, get_node contracted as topology-only, [0.16.0]
  changelog amended, fields sealed pub(crate)), so the footgun class this
  migration fixed is now unrepresentable from outside the engine.
• Engine floor moves to kglite 0.16.0, whose .kgl container v6 is a one-way
  format break: files codingest now writes cannot be read by kglite 0.15.14 or
  earlier (an old reader refuses by version number with FileFormatError
  rather than misreading). The Python requirement widens to
  kglite>=0.16.0,<0.17 — the wheel's compiled writer and the installed
  reader must sit on the same side of the break. v6 also shrinks files
  (upstream: 0.93×/0.79×/0.43× on their three size fixtures).
• The save path uses 0.16.0's kglite::api::io::prepare_kgl_write — the one
  published route for pre-write consolidation, replacing the open-coded
  prepare_save + enable_columnar pair (retired upstream). Also preserves
  copy-on-write lineage across the mutation, which the old spelling did not.

Fixed

• Multi-name from-imports expand per name. from pkg import sub, helper
  previously recorded one bare-module import (the whole statement landed on
  the package, and a call reached only through the submodule name read
  import_backed=false). Each name now records its own import: submodule
  names resolve to their files, symbol names fall back to the module, and
  import_count counts names — uniform with import a, b.
• Docs-pass regression under kglite 0.16.0: MENTIONS/DOCUMENTS edges were
  silently lost. 0.16.0 made every graph columnar from its first node, so
  the docs symbol index — which read node ids/titles via the raw
  DirGraph::get_node route — got Null sentinels, matched nothing, and
  dropped the edges with no error. Reads go through node_view now; every
  frozen parity golden is byte-identical again (verified across all 15
  corpora, goldens NOT regenerated).