codingest 0.2.16

//! Feature-parity regression test for the `codingest` code-tree builder.
//!
//! History: KGLite's in-tree `kglite::code_tree` module was the authority
//! this crate was extracted from. While both existed, `corpus_parity` and
//! `rev_path_parity` built every input with BOTH builders and asserted the
//! graphs were byte-for-byte equivalent. KGLite deleted its in-tree builder
//! on 2026-07-16, so cross-builder comparison is no longer possible — those
//! two tests were removed. The authority they enforced was frozen (while the
//! builders were still verified-identical) into per-corpus SHA-256 golden
//! digests, and the surviving tests carry it forward:
//!
//!   * `golden_parity` builds each corpus `BUILDS_PER_CORPUS` times with
//!     `codingest` and asserts every build's canonical digest matches the
//!     frozen golden (see `tests/goldens/README.md`). Repeating the build is
//!     what makes this the project's **determinism** gate as well: hash
//!     iteration order is randomized per `HashMap` instance, so an
//!     order-dependent builder produces digests that disagree with each other
//!     (reported as NONDETERMINISM) or agree with each other but not the
//!     golden (reported as a behaviour change). This replaced a Makefile-only
//!     step that ran three builds of an *external sibling checkout* and pinned
//!     its exact edge count — a gate whose verdict depended on a repository
//!     this project does not own, which never ran in CI, and which skipped
//!     silently when the checkout was absent.
//!   * `rev_self_consistency` builds the same 2-commit tempdir repo twice with
//!     `codingest` and asserts the graphs are equivalent — INCLUDING the
//!     `revs` / `rev_fp` list properties stamped onto every node and edge — a
//!     determinism check for the multi-rev path (whose digest can't be frozen;
//!     fresh commit SHAs leak into `revs`).
//!
//! `assert_graphs_equiv` compares two graphs along every dimension:
//!   * identical node-type → count maps
//!   * identical edge-type → count maps
//!   * identical sets of (node_type, id)
//!   * identical per-node property maps (full sweep, nothing excluded)
//!   * identical per-edge property maps (full sweep, nothing excluded)
//!
//! Nothing is excluded from the property sweep: file-path properties are
//! stored relative to the project root (`builder/mod.rs`), so they match
//! between two builds of the same tree.

use kglite::api::{DirGraph, GraphRead, Value};
use sha2::{Digest, Sha256};
use std::collections::BTreeMap;
use std::path::{Path, PathBuf};
use std::process::Command;

const CORPORA: &[&str] = &[
    "py_basic",
    "py_inheritance",
    "rust_xfile",
    "ts_callback",
    "cross_ts_py",
    "dup_minified_assets",
    "agc_basic",
    // Added 2026-08-01. Every other corpus is blind to TS/JS import
    // resolution — none contains a single TS `import` statement, so a change
    // to it could land with zero golden movement. Its golden is therefore
    // captured additively (see `tests/goldens/README.md`).
    "ts_monorepo",
    // Added 2026-08-01. No other corpus contains a `const` bound to a
    // function literal, a `function*` in any spelling, or a factory-wrapped
    // binding (`Effect.fn(…)(function*…)`), so depth-0 higher-order-function
    // bindings and the TS grammar vocabulary could be changed — or broken —
    // with zero golden movement. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "ts_hof_binding",
    // Added 2026-08-01. Nothing else in the corpus set declares anything
    // below the top level of a TS file — no nested named binding, no
    // `namespace`, no closure-scoped helper — so the entire scope walk
    // (D1/D2/D3/D4) could be changed, or deleted, with zero golden movement.
    // It pins the shapes that must become nodes *and* the ones that must not
    // (a binding under an anonymous callback, `arr.map(f)` at depth > 0, a
    // plain IIFE), the `#{line}` tie-break, and D3's same-file-only CALLS
    // participation via a cross-file caller that must resolve nothing. Its
    // golden is captured additively (see `tests/goldens/README.md`).
    "ts_closure_scope",
    // Added 2026-08-01. The four committed Python corpora contain nothing but
    // top-level `def`s and plain classes — not one nested definition between
    // them — so the Python scope walk (D1/D2/D3/D4) could be changed, or
    // deleted, with zero golden movement. It pins the shapes that must become
    // nodes (a decorator factory two levels deep, a closure factory, a nested
    // helper, a method-local, a function-local class's methods), the block
    // transparency and lambda rules that are Python's answer to D1 clause 5,
    // the `#{line}` tie-break on both the `if`/`else` and `try`/`except`
    // conditional-definition idioms, and D3 from both sides — a cross-file
    // caller that must resolve nothing against same-file CALLS, REFERENCES_FN
    // and DECORATES edges that must resolve. Its golden is captured additively
    // (see `tests/goldens/README.md`).
    "py_nested_defs",
    // Added 2026-08-01. Not one of the ten pre-existing corpora contains an
    // `.mdx`, an upper-cased `.MD`, or a `.txt` that must stay out — so the
    // docs pass's accepted-extension match could be widened (to `.txt`, the
    // tempting "helpful" change) or narrowed (dropping `.mdx`) with zero
    // golden movement. It pins the `.mdx` arm end to end — frontmatter
    // properties, headings, backtick MENTIONS, a doc→doc DOCUMENTS edge whose
    // target is an `.mdx` and whose source is a `.md`, and a doc→File edge —
    // plus the extension-stripped `:Doc` id (`README.MD` → `README`) and the
    // inertness of embedded JSX/ESM. `NOTES.txt` is markdown-shaped and must
    // contribute nothing. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "docs_mdx",
    // Added 2026-08-10. No other corpus contains two docs in one directory
    // whose names differ only by markup extension, so the concept-id collision
    // policy (`.mdx` > `.md` > `.rst`; the loser is dropped from doc-node
    // emission entirely) could be changed — or regress to the old silent
    // last-write-wins overwrite — with zero golden movement. It pins the
    // survivor's identity (the `.mdx` file's title, `file_path` and mentions),
    // the loser's total absence (a symbol and a link that exist ONLY in
    // `docs/guide.md` must contribute no edge), a link written against the
    // DROPPED spelling (`docs/guide.md`) still resolving to the surviving
    // `docs/guide` node, an uncolliding `.rst` surviving untouched, and a
    // mixed-case `Notes.MD` stripping to `Notes`. Its golden is captured
    // additively (see `tests/goldens/README.md`).
    "docs_ext_collide",
    // Added 2026-08-10. No other corpus registers the same route path from two
    // different files, so Route identity could collapse every methodless
    // `@app.route('/')` in a repo into one node — reporting whichever file the
    // sorted walk reached first as the source location of all of them — with
    // zero golden movement. It pins both sides of the registration model: two
    // files registering `/` are two Route nodes each carrying its own truthful
    // `file_path`/`line_number`, while two registrations of `/dup` inside one
    // file are one registration site with parallel HANDLES edges (the id
    // carries the declaring file, deliberately not the line). Its golden is
    // captured additively (see `tests/goldens/README.md`).
    "py_routes_dup",
    // Added 2026-08-10. The `lang_group` CALLS tier had NO golden coverage:
    // `cross_ts_py` shares no bare name between its TS and Python halves
    // (`createSession` vs `create_session`), and `dup_minified_assets`'s
    // `index.html` contains no `<script>` at all, so no corpus ever reached
    // tier 3 with candidates from two different language families — the whole
    // grouping could be changed, or deleted, with every digest staying green.
    // It pins the case that motivated declaring groups in the registry: an
    // HTML-embedded script (rescoped to `index.html:script_N.<name>`, a qname
    // whose punctuation looks Python-ish) calling a name defined BOTH in a
    // `.js` and in a `.py`, which must resolve to the JavaScript definition.
    // Its golden is captured additively (see `tests/goldens/README.md`).
    "html_js_lang_group",
    // Added 2026-08-15 (post-016 program B0). The four corpora below exist
    // because the import-resolution audit found the defect shapes invisible:
    // rust_xfile carries ONE `pub mod` and zero `use` statements, the Python
    // corpora carry only absolute happy-path imports, and C/C++ and Dart had
    // no files at all — so five verified resolver defects (mcp-servers report,
    // 2026-08-14) had no golden that could see a fix or a regression. Each
    // corpus pins the broken pre-fix behaviour first; the fix phase
    // regenerates its golden with the reason recorded, so the golden diff IS
    // the record of what the fix changed.
    "rust_import", // crate::/super::/self::, use-as aliases, scoped_use_list
    "py_import",   // relative imports, `import a, b`, aliased, from-pkg-import-sub, nested walk
    "cpp_include", // quoted #include vs <system>, same-dir + subdir resolution
    "dart_import", // package: URIs with directory structure; a/x.dart vs b/x.dart
    // Added 2026-08-15 with the R parser (post-016 E2). No other corpus
    // contains a single `.R`/`.r` file, so the whole language — both
    // assignment shapes (`<-`, `=`) and the `\(x)` lambda, S4
    // setClass/setGeneric/setMethod extraction, the `source()` path-import
    // chain A→B→C (`main.R` → `sub/helpers.R` → `sub/deep.R`, including a
    // file-relative `source("deep.R")` from inside `sub/`), the lowercase
    // `.r` extension registration, and the `library(tools)` collision bait
    // (a local `tools.R` that must attract NO edge from the package
    // reference) — could be changed, or deleted, with zero golden movement.
    // Its golden is captured additively (see `tests/goldens/README.md`).
    "r_basic",
    // Added 2026-08-15 (post-016 program E, Julia support). No other corpus
    // contains a `.jl` file, so the entire Julia parser — long-form and
    // short-form (`f(x) = …`) functions, a multi-dispatch pair that the
    // builder's overload pass must keep as two nodes, `module` blocks,
    // structs with fields and `<:` supertypes — could be changed, or deleted,
    // with zero golden movement. It also pins Julia's split import model from
    // both sides: an `include("…")` chain (Main.jl → geometry.jl →
    // shapes/circle.jl) that MUST form File→File edges through the path
    // route, and a `using Downloads` whose name collides with a never-included
    // src/Downloads.jl file bait that must form NO File→File edge — `using`
    // is namespace-shaped and julia is deliberately absent from the
    // file-anchored raw-walk allowlist. Its golden is captured additively
    // (see `tests/goldens/README.md`).
    "julia_basic",
    // Added 2026-08-22. Not one corpus contained a `.cs` file, so the entire
    // C# parser — and with it the ONE call tier that needs a namespace
    // import to fire — could be changed, or deleted, with every digest green.
    // It pins the using-alias fix from BOTH sides, because the pre-fix graph
    // was well-formed and wrong rather than empty: `using_directive` took its
    // first identifier child, so `using Log = MyApp.Logging;` recorded the
    // ALIAS as the imported namespace, and `src/Decoy/Log/Logger.cs` declares
    // a real `namespace Log` holding a real `Logger.Emit` for that mistake to
    // land on. Measured pre-fix on this corpus: `Service.cs` IMPORTS `Log`
    // and `Run` CALLS `Log.Logger.Emit` — a plausible edge onto the decoy, at
    // the same resolution and candidate count as the right one. Post-fix the
    // import lands on `MyApp.Logging` and the call on
    // `MyApp.Logging.Logger.Emit`, still through the `namespace_import` tier,
    // which no other corpus can reach: the tier requires a `.`/`::` after the
    // imported prefix, so PHP's `\`-separated namespaces can never exercise
    // it. `using System` must attract no edge, `Missing.Publish` must resolve
    // to nothing, and the corpus deliberately declares no method with a bare
    // user-defined return type — the second half of the same fix, pinned by
    // unit tests, because the pre-fix parser named such a method after its
    // return type and would have hidden the alias evidence behind a second
    // defect. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "csharp_using_alias",
    // Added 2026-08-22. Not one corpus contained a `.php` file, so PHP's
    // grouped `use` — `use App\Domain\{Billing\Invoice, Catalog\Product};` —
    // had no coverage: `extract_use_imports` never matched the
    // `namespace_use_group` body and recorded ONE import, the bare ancestor
    // `App\Domain`, dropping both members. That is the ancestor-edge family
    // the import walk's `min_end` bound exists to prevent, reintroduced at
    // extraction where the bound cannot see it. Measured pre-fix on this
    // corpus: two IMPORTS edges (`App\Models` and `App\Domain`); post-fix,
    // three (`App\Models`, `App\Domain\Billing`, `App\Domain\Catalog`). The
    // group members are deliberately SUB-NAMESPACE-qualified — a bare
    // `{User, Post}` group trims straight back to the ancestor module and
    // would pin nothing. The corpus also pins the `is_method` inversion from
    // the same commit: top-level `build_report` was stored `is_method=true`
    // and is now `false`, while the four class methods stay `true`. A
    // `use … as Logger` for a package that does not exist must attract no
    // edge, `compute_missing_total` must resolve to nothing, `User
    // implements Identified` pins the IMPLEMENTS edge, and `displayName`
    // — declared on both the interface and the class — pins the `lang_group`
    // fan-out to two candidates. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "php_group_use",
    // Added 2026-08-22. Not one corpus contained a `.java` file, so Java's
    // docstring extraction — `get_doc_comment`, the ONLY producer of a
    // `docstring` property in this parser — could be changed, or deleted,
    // with every digest green. It pins the property from both sides: eight
    // javadoc'd declarations (four types — an interface, an abstract class,
    // a concrete subclass, a final utility class — and one method on each)
    // carry a docstring, and `quiet`, the one method preceded by a plain
    // `//` line comment instead of a javadoc block, must carry NULL. The
    // same `//` line is the other half of the pin: a `// TODO:` inside
    // `AbstractGreeter` must appear in that FILE's `annotations`, which is a
    // different extraction path over the same comment vocabulary. It also
    // pins the namespace import walk's one-segment `min_end` bound with live
    // bait — `import com.acme.Formatter` must form NO edge even though
    // Module `com` exists in the graph (a deeper trim is what landed 252
    // measured false edges on gson's ancestor modules), while
    // `import com.example.util.Text` next to it must resolve. IMPLEMENTS and
    // EXTENDS are pinned across three files, and Java's short
    // `source_type` spelling with them. Its golden is captured additively
    // (see `tests/goldens/README.md`).
    "java_javadoc",
    // Added 2026-08-22. Not one corpus contained a `.go` file, so Go's
    // interface-method arm — `method_elem`, the node tree-sitter-go 0.25
    // renamed `method_spec` to, and the sole producer of interface-method
    // Functions and their HAS_METHOD edges — could be deleted and every
    // digest would stay green. The corpus pins that arm twice over, by two
    // independent detectors: the `Reader.Fetch`/`Reader.Reset` nodes and
    // their two HAS_METHOD edges, and the fan-out of `s.Fetch(…)` in
    // `main`, which must reach BOTH `Reader.Fetch` and `Memory.Fetch` as a
    // two-candidate `lang_group` resolution — if the arm dies, that
    // collapses silently into one `unique_name` edge onto the struct method.
    // `NewMemory` (cross-file, `unique_name`) and `normalize` (called from
    // the sibling file in the same package) pin the ordinary routes;
    // `fmt.Println`, `strings.TrimSpace` and `missingHelper` must resolve to
    // nothing. Two absences are deliberate. Go's implicit interface
    // conformance is unmodeled, so `Memory` satisfying `Reader` produces NO
    // IMPLEMENTS edge. And `import "demo/store"` — the ordinary shape, the
    // module path from `go.mod` plus the package directory — produces NO
    // File→Module IMPORTS edge: `go.mod` is not read (the manifest reader
    // takes only `pyproject.toml` and `Cargo.toml`), and a Go module path is
    // built package-name-FIRST (`store/store` for `store/*.go` in package
    // `store`), so no prefix of any real import specifier can ever match
    // one. Both absences are pinned so that a fix has to move this digest.
    // Its golden is captured additively (see `tests/goldens/README.md`).
    "go_interface",
    // Added 2026-08-22. Not one corpus contained a `.swift` file, so the
    // whole Swift parser was invisible to the golden net. It pins the
    // `is_method` inversion fixed alongside PHP's: top-level `trim` was
    // stored `is_method=true` (the parser's own comment asserted the
    // opposite of its code) and is now `false`, while the struct's and the
    // enum's methods stay `true`. `Package.swift` is REAL PARSED SOURCE
    // here, not an inert manifest — it pins the non-SwiftPM fallback module
    // path `swift_basic.Package`, derived from the corpus directory name,
    // and `import PackageDescription` inside it must attract no edge.
    // Swift is one of only two languages in the set with an implicit module
    // hierarchy, so the `Demo`/`Support` parents and their HAS_SUBMODULE
    // chains are pinned here too, together with `Greeter.swift`'s
    // `import Support` resolving onto the `Support` Module. Two absences are
    // deliberate and are what a future fix would move: `struct Greeter:
    // Greeting` produces NO IMPLEMENTS edge (the inheritance specifier is
    // unparsed, as the parser's header admits), and `Text.shout` is reached
    // by its terminal segment alone because Swift call extraction never
    // keeps the receiver. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "swift_basic",
    // Added 2026-08-22. `dart_import` is the only other Dart corpus and it
    // contains no `part` file at all, so Dart's `part of` handling — the one
    // place a file's module path comes from something other than its own
    // location on disk — had zero golden coverage and could be changed, or
    // broken, with every digest staying green. It pins the property the
    // feature exists for: a library split across `lib/collection.dart` and
    // two parts under `lib/src/` is ONE module, with all three files hanging
    // off `dart_part_of.lib.collection`. The pre-fix derivation kept only the
    // URI's stem (`{pkg}.{stem}`), so the parts landed in a phantom
    // `dart_part_of.collection` module that no file declared and the parent
    // library never joined; the `../collection.dart` spelling is what makes
    // the dropped directory segments visible. `fromA` calling the parent
    // file's `seed` pins that the cross-file CALLS edge still resolves once
    // the three files share a module. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "dart_part_of",
    // Added 2026-08-22. Every Rust `use` in every other corpus sits at file
    // level, so the scope a `use` is written in never mattered and the flat,
    // scope-less `FileInfo::imports` list could stay wrong indefinitely: a
    // `use super::…` inside an inline `mod` was popped against the FILE's
    // module, one level too shallow, and landed on the parent module's file.
    // `mod tests { use super::*; }` — the most common shape in Rust source —
    // is exactly that bug. The corpus pins all three arms of the re-anchoring
    // (`parsers::rust_lang::RustParser::rebase_inline_mod_use`): supers equal
    // to the inline depth resolve to the file itself and form NO edge (the
    // self-guard eats them), a surplus super still pops the file's real
    // parent and reaches `beta.rs`, and two nested inline levels cancel two
    // supers. `beta.rs`'s file-level `use crate::…` is the control that must
    // not move. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "rust_inline_mod",
    // Added 2026-08-22. The four committed Python corpora all put their
    // packages where the importing file can already see them, so the ONE
    // layout `pyproject.toml` made standard — the package under `src/`,
    // imported by its absolute installed name from a sibling `tests/` tree —
    // had no golden at all: `from pkg.util import helper` in `tests/`
    // resolved to nothing, and it could have kept resolving to nothing
    // forever with every digest green. It pins the evidence-gated src root
    // from both ends: the import DOES resolve (File→File and File→Module,
    // onto `src/pkg/util.py`), and the root is offered only because the file
    // set contains a `src/` directory — `py_import`, which has none, is the
    // control that must not move. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "py_src_layout",
    // Added 2026-08-22. `cpp_include`'s every `#include` sits at a file's top
    // level, so the single most common real-world header prologue —
    // `#ifdef __cplusplus` / `extern "C" {` / `#endif` — had no coverage, and
    // the includes inside it were silently dropped (measured on
    // DaveGamble/cJSON: 5 of 96 quoted includes missed, `cJSON_Utils.h` →
    // `cJSON.h` among them — the library's own core edge). It pins both ways
    // tree-sitter renders that prologue: `utils.h` closes the block, so the
    // body becomes a `linkage_specification` the C router had no arm for;
    // `decls_begin.h` leaves it open, so the rest of the file collapses into
    // one `ERROR` recovery subtree. It also pins what must NOT be extracted —
    // the `<vector.h>` angle include inside the same region (a REAL project
    // file, excluded by its delimiters) and `phantom.h`, named by an
    // include-SHAPED line with no `#`, which only a text scan of the ERROR
    // region could ever reach. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "cpp_extern_c",
    // Added 2026-08-22. No other corpus contains a leading-`/` web reference
    // at all, so the whole served-root question was invisible: a built site
    // under `dist/` is served with `dist/` as `/`, and `/_astro/app.css`
    // resolved against the project root only — never reaching
    // `dist/_astro/app.css`, which is right there in the graph. It pins the
    // new candidate AND its precedence: `/shared/reset.css`, which resolves
    // at the project root today, must keep resolving there, so the project
    // root is still tried first. Its golden is captured additively (see
    // `tests/goldens/README.md`).
    "web_served_root",
];

/// Independent builds of each corpus per `golden_parity` run.
///
/// One build proves the output still matches the frozen golden; repeating it
/// also proves the output does not depend on hash iteration order. The
/// original nondeterminism bug (randomized `HashMap` iteration over DEFINES
/// (source_type, target_type) pairs — whichever pair went first got
/// skip-existence-check dedup semantics) is reproduced by the
/// `dup_minified_assets` corpus, which defines the same selector/element id
/// from one file more than once. Three builds is enough to make an
/// order-dependent builder fail loudly and quickly; the digest comparison
/// against the golden catches the residual case where all three happen to
/// agree on a wrong order.
const BUILDS_PER_CORPUS: usize = 3;

fn corpus_root() -> PathBuf {
    // tests/parity.rs lives in crates/code-tree; corpus is at the workspace root.
    Path::new(env!("CARGO_MANIFEST_DIR"))
        .join("..")
        .join("..")
        .join("tests")
        .join("corpus")
}

/// Deterministic canonical string for a property value. `{:?}` on `Value`
/// is stable (enums, not maps) and captures scalars and nested lists alike.
fn canon(v: &Value) -> String {
    format!("{v:?}")
}

/// node_type -> count.
fn node_type_counts(g: &DirGraph) -> BTreeMap<String, usize> {
    let mut m = BTreeMap::new();
    for idx in g.graph.node_indices() {
        if let Some(n) = g.node_view(idx) {
            *m.entry(n.node_type_str(&g.interner).to_string())
                .or_insert(0) += 1;
        }
    }
    m
}

/// connection_type -> count.
fn edge_type_counts(g: &DirGraph) -> BTreeMap<String, usize> {
    let mut m = BTreeMap::new();
    for e in g.graph.edge_indices() {
        if let Some(edge) = g.graph.edge_weight(e) {
            *m.entry(edge.connection_type_str(&g.interner).to_string())
                .or_insert(0) += 1;
        }
    }
    m
}

/// Sorted (node_type, id) identity set.
fn node_identities(g: &DirGraph) -> Vec<(String, String)> {
    // Read through `node_view`, not the raw `node_weight`. kglite 0.16.0 made
    // every graph columnar from its first node and deleted the row layout, so a
    // node's own struct no longer carries its id — `NodeData::id()` yields the
    // `Null` sentinel and this digest silently became a list of "NULL". The
    // view resolves the sentinel through the type's column store.
    let mut v: Vec<(String, String)> = g
        .graph
        .node_indices()
        .filter_map(|i| g.node_view(i))
        .map(|n| (n.node_type_str(&g.interner).to_string(), n.id().to_string()))
        .collect();
    v.sort();
    v
}

/// Full per-node property sweep: sorted list of
/// (node_type, id, sorted{prop -> canon}). Nothing excluded.
fn node_props(g: &DirGraph) -> Vec<(String, String, BTreeMap<String, String>)> {
    let mut v: Vec<(String, String, BTreeMap<String, String>)> = g
        .graph
        .node_indices()
        .filter_map(|i| g.node_view(i))
        .map(|n| {
            let props: BTreeMap<String, String> = n
                .properties_cloned(&g.interner)
                .iter()
                .map(|(k, val)| (k.clone(), canon(val)))
                .collect();
            (
                n.node_type_str(&g.interner).to_string(),
                n.id().to_string(),
                props,
            )
        })
        .collect();
    v.sort();
    v
}

/// Full per-edge property sweep: sorted list of
/// (conn, src_id, tgt_id, sorted{prop -> canon}). Edges can share an
/// identity triple, so this is a multiset kept as a sorted Vec.
fn edge_props(g: &DirGraph) -> Vec<(String, String, String, BTreeMap<String, String>)> {
    let mut v: Vec<(String, String, String, BTreeMap<String, String>)> = g
        .graph
        .edge_indices()
        .filter_map(|e| {
            let edge = g.graph.edge_weight(e)?;
            let (s, t) = g.graph.edge_endpoints(e)?;
            // Endpoint ids come from the column store — see `node_identities`.
            let sn = g.node_view(s)?;
            let tn = g.node_view(t)?;
            let props: BTreeMap<String, String> = edge
                .properties_cloned(&g.interner)
                .iter()
                .map(|(k, val)| (k.clone(), canon(val)))
                .collect();
            Some((
                edge.connection_type_str(&g.interner).to_string(),
                sn.id().to_string(),
                tn.id().to_string(),
                props,
            ))
        })
        .collect();
    v.sort();
    v
}

// ── canonical rendering + golden digest ──────────────────────────────────
//
// The golden oracle renders a graph to a single deterministic string built
// from the *exact same* data the equivalence assertions sweep — node-type
// counts, edge-type counts, the sorted (node_type, id) identity set, the full
// per-node property sweep, and the full per-edge property sweep — then hashes
// it with SHA-256. Because every input is already sorted (BTreeMap / sorted
// Vec) and `canon` is `Value`'s stable `Debug` form, the rendering is
// reproducible across runs and machines. Nothing is excluded: whatever
// `assert_graphs_equiv` compares, the digest covers.

/// Deterministic, exhaustive canonical rendering of a graph. Two graphs render
/// to the same string iff `assert_graphs_equiv` would consider them equivalent.
fn canonical_graph_string(g: &DirGraph) -> String {
    let mut s = String::new();

    s.push_str("## node_type_counts\n");
    for (ty, n) in node_type_counts(g) {
        s.push_str(&format!("{ty}\t{n}\n"));
    }

    s.push_str("## edge_type_counts\n");
    for (ty, n) in edge_type_counts(g) {
        s.push_str(&format!("{ty}\t{n}\n"));
    }

    s.push_str("## node_identities\n");
    for (ty, id) in node_identities(g) {
        s.push_str(&format!("{ty}\t{id}\n"));
    }

    s.push_str("## node_props\n");
    for (ty, id, props) in node_props(g) {
        s.push_str(&format!("{ty}\t{id}\n"));
        for (k, v) in props {
            s.push_str(&format!("\t{k}={v}\n"));
        }
    }

    s.push_str("## edge_props\n");
    for (conn, src, tgt, props) in edge_props(g) {
        s.push_str(&format!("{conn}\t{src}\t{tgt}\n"));
        for (k, v) in props {
            s.push_str(&format!("\t{k}={v}\n"));
        }
    }

    s
}

/// SHA-256 (lowercase hex) of the canonical graph rendering — the golden digest.
fn graph_digest(g: &DirGraph) -> String {
    let mut hasher = Sha256::new();
    hasher.update(canonical_graph_string(g).as_bytes());
    let out = hasher.finalize();
    out.iter().map(|b| format!("{b:02x}")).collect()
}

fn goldens_dir() -> PathBuf {
    Path::new(env!("CARGO_MANIFEST_DIR"))
        .join("tests")
        .join("goldens")
}

/// Read the stored golden digest for a corpus (one hex line), trimmed.
fn read_golden(corpus: &str) -> String {
    let path = goldens_dir().join(format!("{corpus}.sha256"));
    std::fs::read_to_string(&path)
        .unwrap_or_else(|e| {
            panic!(
                "missing golden {} — capture with `cargo test -p codingest --test parity -- --ignored capture_goldens`: {e}",
                path.display()
            )
        })
        .trim()
        .to_string()
}

/// Assert the two graphs are equivalent along every dimension, with
/// pinpointed diffs on failure.
fn assert_graphs_equiv(label: &str, left: &DirGraph, right: &DirGraph) {
    // 1. node-type counts
    let (a, b) = (node_type_counts(left), node_type_counts(right));
    assert_eq!(
        a, b,
        "[{label}] node-type counts differ\nleft={a:?}\nright={b:?}"
    );

    // 2. edge-type counts
    let (a, b) = (edge_type_counts(left), edge_type_counts(right));
    assert_eq!(
        a, b,
        "[{label}] edge-type counts differ\nleft={a:?}\nright={b:?}"
    );

    // 3. node identity set
    let (a, b) = (node_identities(left), node_identities(right));
    if a != b {
        let only_in: Vec<_> = a.iter().filter(|x| !b.contains(x)).take(10).collect();
        let only_st: Vec<_> = b.iter().filter(|x| !a.contains(x)).take(10).collect();
        panic!(
            "[{label}] node identity sets differ\nonly left (<=10): {only_in:?}\nonly right (<=10): {only_st:?}"
        );
    }

    // 4. per-node property sweep
    let (a, b) = (node_props(left), node_props(right));
    if a != b {
        for (x, y) in a.iter().zip(b.iter()) {
            if x != y {
                panic!(
                    "[{label}] node properties differ\nfirst mismatch:\n  left=({}, {}) {:?}\n  right=({}, {}) {:?}",
                    x.0, x.1, x.2, y.0, y.1, y.2
                );
            }
        }
        panic!(
            "[{label}] node property lists differ in length: left={} right={}",
            a.len(),
            b.len()
        );
    }

    // 5. per-edge property sweep
    let (a, b) = (edge_props(left), edge_props(right));
    if a != b {
        for (x, y) in a.iter().zip(b.iter()) {
            if x != y {
                panic!(
                    "[{label}] edge properties differ\nfirst mismatch:\n  left=({}, {}->{}) {:?}\n  right=({}, {}->{}) {:?}",
                    x.0, x.1, x.2, x.3, y.0, y.1, y.2, y.3
                );
            }
        }
        panic!(
            "[{label}] edge property lists differ in length: left={} right={}",
            a.len(),
            b.len()
        );
    }
}

// ── golden parity (the authority carried forward past the in-tree deletion) ─
//
// Builds every corpus with ONLY the `codingest` builder, digests the result,
// and compares it to the frozen golden captured from the in-tree authority
// (see `tests/goldens/README.md`). This is the test that carries the
// authority forward now that KGLite has deleted its in-tree builder: it keeps
// proving that codingest reproduces the frozen graph byte-for-byte.
//
// It is also the determinism gate: each corpus is built `BUILDS_PER_CORPUS`
// times and every build must produce the same digest as every other *and* as
// the golden. The two failure modes are reported separately, because they call
// for opposite responses — a behaviour change may be legitimate and
// regenerated, while nondeterminism is always a bug.
#[test]
fn golden_parity() {
    let root = corpus_root();
    for name in CORPORA {
        let dir = root.join(name);
        assert!(dir.is_dir(), "missing corpus dir: {}", dir.display());

        let build = || {
            codingest::builder::run_with_options(&dir, false, true, None, None, true)
                .unwrap_or_else(|e| panic!("[{name}] codingest build failed: {e}"))
        };

        let first = build();
        let got = graph_digest(&first);

        // 1. Determinism: repeat builds must be identical to the first.
        for run in 2..=BUILDS_PER_CORPUS {
            let again = build();
            let again_digest = graph_digest(&again);
            if again_digest != got {
                // Pinpoint the divergence before reporting it.
                assert_graphs_equiv(
                    &format!("{name} nondeterminism run 1 vs {run}"),
                    &first,
                    &again,
                );
                panic!(
                    "[{name}] NONDETERMINISM: build 1 and build {run} of the same corpus \
                     produced different digests ({got} != {again_digest}) but compared equal \
                     dimension-by-dimension — widen the canonical rendering."
                );
            }
        }

        // 2. Frozen authority: the (now proven stable) digest must be the golden.
        let want = read_golden(name);
        assert_eq!(
            got, want,
            "[{name}] golden digest mismatch\n  golden (frozen authority) = {want}\n  codingest build           = {got}\n\
             All {BUILDS_PER_CORPUS} builds agreed, so this is a builder BEHAVIOUR change, not nondeterminism.\n\
             If it is deliberate, regenerate with\n\
             `cargo test -p codingest --test parity -- --ignored capture_goldens`."
        );
    }
}

/// Regeneration path for the corpus goldens. `#[ignore]` so it never runs in
/// the normal suite — invoke explicitly:
///
///   cargo test -p codingest --test parity -- --ignored capture_goldens
///
/// It writes `tests/goldens/<corpus>.sha256` for every corpus.
///
/// AUTHORITY NOTE: KGLite deleted its in-tree builder on 2026-07-16, so
/// regeneration now captures from the `codingest` builder — codingest is its
/// own oracle. The frozen digests it overwrites were originally captured (on
/// 2026-07-16) from the last in-sync in-tree authority, so only regenerate
/// when a deliberate builder-behavior change makes a golden legitimately
/// stale (see `tests/goldens/README.md`), never to paper over a red
/// `golden_parity`.
#[test]
#[ignore = "regeneration path; run explicitly with --ignored capture_goldens"]
fn capture_goldens() {
    let root = corpus_root();
    let out_dir = goldens_dir();
    std::fs::create_dir_all(&out_dir).expect("create goldens dir");
    for name in CORPORA {
        let dir = root.join(name);
        assert!(dir.is_dir(), "missing corpus dir: {}", dir.display());

        // Capture from the codingest builder — the authority, now that the
        // in-tree builder is gone.
        let g = codingest::builder::run_with_options(&dir, false, true, None, None, true)
            .unwrap_or_else(|e| panic!("[{name}] codingest build failed: {e}"));
        let digest = graph_digest(&g);
        let path = out_dir.join(format!("{name}.sha256"));
        std::fs::write(&path, format!("{digest}\n")).expect("write golden");
        eprintln!("captured {name} -> {digest}");
    }
}

// ── serialized-bytes determinism ─────────────────────────────────────────
//
// `golden_parity` digests the graph as it exists IN MEMORY, from sorted maps
// and sorted vectors. That rendering is structurally insulated from the order
// properties were inserted in — which is precisely the bug class that bit this
// project before: two builds with identical in-memory digests wrote `.kgl`
// files that differed byte-for-byte, because edge-property insertion order
// leaked into the serialization (fixed engine-side in kglite 0.14.5; the
// codingest-side assertion never existed). CALLS edges now carry three more
// properties each, so the exposure is larger, not smaller.
//
// This is the net the in-memory digest cannot provide: build the same corpus
// repeatedly, save each build, compare the FILES.
#[test]
fn reloaded_graph_renders_identically() {
    // The property kglite 0.16.0's headline change is about: a freshly built
    // graph and a reloaded one are "the same shape". Assert it at CONTENT
    // level — build → save `.kgl` (v6) → load → the canonical rendering of the
    // reloaded graph is byte-identical to the in-memory one. `kgl_bytes`
    // proves the WRITER is deterministic without ever loading a file; the cli
    // query tests reload without digesting. Until this test, "a reload loses
    // or rewrites values" was caught by nothing — and a reloaded graph is
    // exactly where the columnar sentinel lived on 0.15.x.
    let tmp = tempfile::Builder::new()
        .prefix("codingest-reload-digest-")
        .tempdir()
        .unwrap();
    let root = corpus_root();
    // Densest edge-property corpus + a docs corpus (Doc nodes carry the widest
    // string properties) + the sparse-column AGC shape.
    for name in ["ts_monorepo", "docs_mdx", "agc_basic"] {
        let dir = root.join(name);
        assert!(dir.is_dir(), "missing corpus dir: {}", dir.display());
        let dest = tmp.path().join(format!("{name}.kgl"));
        let built =
            codingest::builder::run_with_options(&dir, false, true, Some(&dest), None, true)
                .unwrap_or_else(|e| panic!("[{name}] build failed: {e}"));
        let reloaded = kglite::api::io::load_file(&dest.to_string_lossy())
            .unwrap_or_else(|e| panic!("[{name}] reload failed: {e}"));
        assert_eq!(
            canonical_graph_string(&built),
            canonical_graph_string(&reloaded),
            "[{name}] reloaded graph renders differently from the one saved"
        );
    }
}

#[test]
fn kgl_bytes_are_stable_across_builds() {
    let tmp = tempfile::Builder::new()
        .prefix("codingest-kgl-bytes-")
        .tempdir()
        .unwrap();
    let root = corpus_root();

    // `ts_monorepo` carries the densest multi-property CALLS edge set
    // (resolution + candidates + import_backed on every edge); `agc_basic`
    // exercises the sparse-column path, where the semantic pass leaves those
    // three null while populating raw_targets / via / address_lines.
    for name in ["ts_monorepo", "agc_basic"] {
        let dir = root.join(name);
        assert!(dir.is_dir(), "missing corpus dir: {}", dir.display());

        let mut bytes: Vec<Vec<u8>> = Vec::new();
        for run in 1..=BUILDS_PER_CORPUS {
            let dest = tmp.path().join(format!("{name}-{run}.kgl"));
            codingest::builder::run_with_options(&dir, false, true, Some(&dest), None, true)
                .unwrap_or_else(|e| panic!("[{name}] build {run} failed: {e}"));
            bytes.push(std::fs::read(&dest).expect("read written .kgl"));
        }

        for run in 2..=BUILDS_PER_CORPUS {
            let (first, again) = (&bytes[0], &bytes[run - 1]);
            if first == again {
                continue;
            }
            let at = first
                .iter()
                .zip(again.iter())
                .position(|(a, b)| a != b)
                .map(|i| i.to_string())
                .unwrap_or_else(|| "end (length differs)".into());
            panic!(
                "[{name}] .kgl BYTES differ between build 1 and build {run} \
                 ({} vs {} bytes, first difference at offset {at}). The in-memory \
                 parity digest cannot see this: it is built from sorted maps, so \
                 property INSERTION order is invisible to it while remaining \
                 visible in the serialized file.",
                first.len(),
                again.len()
            );
        }
    }
}

/// Diagnostic dump of every corpus's canonical rendering, one file per
/// corpus, into the directory named by `CODINGEST_CANONICAL_DUMP`.
///
/// This is the tool the golden-regeneration protocol runs on: when a change is
/// *supposed* to move only one section of the rendering (e.g. adding edge
/// properties must not touch `node_type_counts`, `edge_type_counts`,
/// `node_identities` or `node_props`), dumping before and after and diffing
/// the sections proves it. A digest tells you *that* something moved; only the
/// rendering tells you *what*, and blessing an edge-set change inside a
/// "properties-only" regeneration would hide it forever.
///
///   CODINGEST_CANONICAL_DUMP=/tmp/before \
///     cargo test -p codingest --test parity -- --ignored dump_canonical
#[test]
#[ignore = "diagnostic; run explicitly with --ignored dump_canonical"]
fn dump_canonical() {
    let out_dir = PathBuf::from(
        std::env::var("CODINGEST_CANONICAL_DUMP")
            .expect("set CODINGEST_CANONICAL_DUMP to the output directory"),
    );
    std::fs::create_dir_all(&out_dir).expect("create dump dir");
    let root = corpus_root();
    for name in CORPORA {
        let dir = root.join(name);
        let g = codingest::builder::run_with_options(&dir, false, true, None, None, true)
            .unwrap_or_else(|e| panic!("[{name}] codingest build failed: {e}"));
        let path = out_dir.join(format!("{name}.txt"));
        std::fs::write(&path, canonical_graph_string(&g)).expect("write dump");
        eprintln!("dumped {name} -> {}", path.display());
    }
}

// ── rev-path parity ──────────────────────────────────────────────────────

fn git_available() -> bool {
    Command::new("git")
        .arg("--version")
        .output()
        .map(|o| o.status.success())
        .unwrap_or(false)
}

fn git(dir: &Path, args: &[&str]) -> String {
    let out = Command::new("git")
        .arg("-C")
        .arg(dir)
        .args(args)
        .output()
        .expect("git spawn");
    assert!(
        out.status.success(),
        "git {args:?} failed: {}",
        String::from_utf8_lossy(&out.stderr)
    );
    String::from_utf8_lossy(&out.stdout).trim().to_string()
}

fn git_commit(dir: &Path, msg: &str) -> String {
    git(dir, &["add", "-A"]);
    git(
        dir,
        &[
            "-c",
            "user.email=t@t",
            "-c",
            "user.name=t",
            "commit",
            "-q",
            "-m",
            msg,
        ],
    );
    git(dir, &["rev-parse", "HEAD"])
}

/// Build a throwaway 2-commit python repo in `repo` and return the two commit
/// SHAs (oldest → newest). rev2 widens `foo`'s signature (a fingerprint
/// change), drops `gone`, adds `bar`, and makes `driver` call `bar` too.
fn build_two_commit_repo(repo: &Path) -> Vec<String> {
    git(repo, &["init", "-q"]);

    // rev1: a small python package with two functions and a caller.
    std::fs::create_dir_all(repo.join("pkg")).unwrap();
    std::fs::write(
        repo.join("pkg/mod_a.py"),
        "def foo(a):\n    return a + 1\n\n\ndef gone():\n    return 0\n",
    )
    .unwrap();
    std::fs::write(
        repo.join("pkg/mod_b.py"),
        "from pkg.mod_a import foo\n\n\ndef driver(x):\n    return foo(x)\n",
    )
    .unwrap();
    let s1 = git_commit(repo, "rev1");

    // rev2: widen foo's signature (fingerprint change), drop `gone`, add `bar`,
    // and make driver call the new bar too.
    std::fs::write(
        repo.join("pkg/mod_a.py"),
        "def foo(a, b):\n    return a + b\n\n\ndef bar(y):\n    return y * 2\n",
    )
    .unwrap();
    std::fs::write(
        repo.join("pkg/mod_b.py"),
        "from pkg.mod_a import foo, bar\n\n\ndef driver(x):\n    return foo(x, 1) + bar(x)\n",
    )
    .unwrap();
    let s2 = git_commit(repo, "rev2");

    vec![s1, s2]
}

fn assert_has_stamped_revs(g: &DirGraph) {
    let has_revs = g
        .graph
        .node_indices()
        .filter_map(|i| g.node_view(i))
        .any(|n| matches!(n.get_property_value("revs"), Some(Value::List(_))));
    assert!(
        has_revs,
        "expected stamped `revs` list properties on rev-merge graph"
    );
}

// ── rev self-consistency (the multi-rev survivor) ──────────────────────────
//
// The rev fixture can't be frozen into a golden (commit SHAs leak into the
// `revs` property, so its digest is unstable across from-scratch runs — see
// `tests/goldens/README.md`). Instead we prove the codingest builder is
// self-consistent: building the SAME two-commit repo twice (identical SHAs,
// since it's one repo) yields two equivalent graphs, INCLUDING the stamped
// `revs`/`rev_fp` provenance.
#[test]
fn rev_self_consistency() {
    if !git_available() {
        eprintln!("SKIP rev_self_consistency: git not available in test environment");
        return;
    }

    let tmp = tempfile::Builder::new()
        .prefix("codingest-rev-selfconsistency-")
        .tempdir()
        .unwrap();
    let repo = tmp.path();
    let revs = build_two_commit_repo(repo);

    let build = || {
        codingest::rev::build_code_tree_revs(repo, &revs, Some(repo), false, true, None, None, true)
            .expect("codingest build_code_tree_revs")
    };
    let first = build();
    let second = build();

    // Confirm the multi-rev provenance is actually present before comparing.
    assert_has_stamped_revs(&first);

    // Two builds of the same fixed rev input must be byte-for-byte equivalent.
    assert_graphs_equiv("rev-self-consistency", &first, &second);
}