@event4u/agent-config 12.0.0 → 13.0.0

package/src/scripts/_lib/bench_ab_complexity.ts

@@ -0,0 +1,439 @@
+/**
+ * Cognitive complexity per function — the T2 endpoint of the Phase-3 metric pair
+ * (`internal/bench/ab-v2-phase3-PREREG.md`), S0.3 delta #11.
+ *
+ * WHY THIS EXISTS. The size claim is a **pair**, never a single number: an arm
+ * may only claim a size win if median added lines fell *without* median
+ * cognitive complexity per changed function rising. Lines down and complexity up
+ * is golfing, and a lines-only report actively rewards it. Until this endpoint
+ * existed, T1 could not be evaluated either — half a pair is no result, not a
+ * partial one — which is why a granted spend ceiling sat unspent.
+ *
+ * WHAT IT COMPUTES. Cognitive complexity as published by G. Ann Campbell
+ * (SonarSource, 2018): a score that rises with how hard control flow is to
+ * *follow*, not with how many paths exist. Three rules, applied to a
+ * tree-sitter AST:
+ *
+ *   1. **Increment** for each break in the linear flow — `if`, `else if`,
+ *      `else`, `for`, `foreach`, `while`, `do`, `catch`, `switch`, the ternary,
+ *      and a `break`/`continue` that targets a label.
+ *   2. **Nesting penalty**: a structure from rule 1 that *nests* costs its own
+ *      +1 plus the current nesting depth. `else`/`else if` take the flat +1 and
+ *      never a nesting penalty, because they add no new indentation to follow.
+ *   3. **Sequences of binary logical operators** cost +1 per *sequence*, not per
+ *      operator: `a && b && c` is one increment, `a && b || c` is two.
+ *
+ * WHY NOT A MATURE TOOL — the ladder, walked rather than asserted. The roadmap
+ * text says complexity "is computed deterministically per stack by mature
+ * tooling", so the alternatives were checked in the order
+ * [`improve-before-implement`](../../rules/improve-before-implement.md) § ladder
+ * requires. Reuse-in-repo: nothing computes complexity anywhere in the tree.
+ * Installed dependency: `eslint` ships a `complexity` rule, but that is
+ * **cyclomatic**, which is the metric F9 explicitly rejects — it scores a flat
+ * `switch` above a triply-nested `if`, so it cannot detect golfing.
+ * `eslint-plugin-sonarjs` implements the real metric and is **not installed**,
+ * covers JS/TS only, and would leave the corpus's PHP fixtures unmeasured. The
+ * ABI-pinned tree-sitter pair already in this repo parses all three languages
+ * the corpus actually contains, and `code_graph/loader.ts` already wraps it. So
+ * this module is a **new implementation over an existing parser**, not a new
+ * dependency — the smallest rung that carries the requirement.
+ *
+ * TWO DEVIATIONS FROM THE PUBLISHED METRIC, NAMED RATHER THAN BURIED. Both are
+ * deliberate and neither favours a treatment arm, because the metric is only
+ * ever compared against itself across arms:
+ *
+ *   - **No recursion increment.** Campbell adds +1 for a recursive call. That
+ *     needs name resolution against a call graph; the fragility it would buy is
+ *     not worth one point on a metric used only for a paired comparison.
+ *   - **A nested function is its own unit.** Campbell attributes a nested
+ *     function's score to its enclosing method. Here every function, method,
+ *     arrow and closure is scored separately and enters the median as its own
+ *     observation. The endpoint is defined as complexity *per function*, so
+ *     rolling a closure into its parent would make one long function with three
+ *     small closures indistinguishable from one dense one.
+ *
+ * UNSUPPORTED FILES ARE REPORTED, NEVER SCORED ZERO. A `.py` or `.sql` fixture
+ * has no grammar here; it lands in `unsupported_files` and contributes no
+ * observation. A zero would read as "this file is simple", which is a different
+ * claim from "we cannot measure it" — the same distinction the cost sheet draws
+ * between an unpriceable model and a free one.
+ */
+import * as fs from 'node:fs';
+import * as path from 'node:path';
+
+import { getParser, type TsNode } from '../code_graph/loader.js';
+import { EXT_LANG, type Lang } from '../code_graph/types.js';
+
+export interface FunctionComplexity {
+    /** Declared name, or `<anon>` for an unnamed closure/arrow. */
+    name: string;
+    /** 1-based line of the function's first token. */
+    line: number;
+    complexity: number;
+}
+
+export interface FileComplexity {
+    path: string;
+    /** `null` when no grammar covers the extension — then `functions` is empty. */
+    lang: Lang | null;
+    functions: FunctionComplexity[];
+}
+
+export interface ComplexityRollup {
+    /** Median complexity across every scored function; `null` when none were scored. */
+    median: number | null;
+    /** How many function units the median is over. */
+    n_functions: number;
+    /** Changed files whose extension no grammar covers — measured as nothing, not as zero. */
+    unsupported_files: string[];
+    /** Changed files that were missing on disk (deleted by the run). */
+    missing_files: string[];
+    per_file: FileComplexity[];
+}
+
+// ── node vocabulary ────────────────────────────────────────────────────────
+//
+// Every set below was read off a real parse (probe over the three grammars),
+// never inferred from the language spec — a node type that does not exist in the
+// grammar silently scores nothing, which is exactly the defect an unverified
+// list produces.
+
+/** A scored unit. Each contributes one observation to the median. */
+const FUNCTION_NODES: Record<Lang, ReadonlySet<string>> = {
+    javascript: new Set([
+        'function_declaration',
+        'function_expression',
+        'generator_function',
+        'generator_function_declaration',
+        'arrow_function',
+        'method_definition',
+    ]),
+    typescript: new Set([
+        'function_declaration',
+        'function_expression',
+        'generator_function',
+        'generator_function_declaration',
+        'arrow_function',
+        'method_definition',
+    ]),
+    php: new Set([
+        'function_definition',
+        'method_declaration',
+        'anonymous_function_creation_expression',
+        'anonymous_function',
+        'arrow_function',
+    ]),
+};
+
+/** +1 **and** a nesting penalty, and raises nesting for everything inside. */
+const NESTING_NODES: Record<Lang, ReadonlySet<string>> = {
+    javascript: new Set([
+        'if_statement',
+        'for_statement',
+        'for_in_statement',
+        'while_statement',
+        'do_statement',
+        'switch_statement',
+        'catch_clause',
+        'ternary_expression',
+    ]),
+    typescript: new Set([
+        'if_statement',
+        'for_statement',
+        'for_in_statement',
+        'while_statement',
+        'do_statement',
+        'switch_statement',
+        'catch_clause',
+        'ternary_expression',
+    ]),
+    php: new Set([
+        'if_statement',
+        'for_statement',
+        'foreach_statement',
+        'while_statement',
+        'do_statement',
+        'switch_statement',
+        // PHP 8 `match` is `switch`'s expression twin and must score like it.
+        // It was missing at first review and scored 0 against its `switch`
+        // twin's 1 — a silent zero pointing in exactly the golfing direction
+        // this endpoint exists to catch, since rewriting a `switch` as a
+        // `match` is a textbook line-saving transform.
+        'match_expression',
+        'catch_clause',
+        'conditional_expression',
+    ]),
+};
+
+/** +1 flat, no nesting penalty and no nesting increase (`else`, `else if`). */
+const FLAT_NODES: Record<Lang, ReadonlySet<string>> = {
+    javascript: new Set(['else_clause']),
+    typescript: new Set(['else_clause']),
+    php: new Set(['else_clause', 'else_if_clause']),
+};
+
+const LOGICAL_OPERATORS: Record<Lang, ReadonlySet<string>> = {
+    javascript: new Set(['&&', '||', '??']),
+    typescript: new Set(['&&', '||', '??']),
+    php: new Set(['&&', '||', '??', 'and', 'or', 'xor']),
+};
+
+const BINARY_NODES = new Set(['binary_expression']);
+
+/** `break outer;` / `continue outer;` — a jump the reader must chase. */
+const LABELLED_JUMP_NODES = new Set(['break_statement', 'continue_statement']);
+
+// ── traversal helpers ──────────────────────────────────────────────────────
+
+function namedChildren(n: TsNode): TsNode[] {
+    const out: TsNode[] = [];
+    for (let i = 0; i < n.namedChildCount; i += 1) {
+        const c = n.namedChild(i);
+        if (c) out.push(c);
+    }
+    return out;
+}
+
+/**
+ * The operator token of a binary expression.
+ *
+ * `childForFieldName('operator')` is the accessor all three grammars answer, so
+ * it is the primary path. The fallback scans the children for a token whose type
+ * IS one of the operators we care about — an operator token's `type` is the
+ * literal (`&&`, `||`, `??`), which is what makes this decidable without asking
+ * whether a node is named.
+ *
+ * An earlier version fell back to "the first unnamed child", identified by
+ * comparing node objects with `===`. That was wrong twice over: the WASM binding
+ * re-creates a wrapper on every accessor call, so the identity test could never
+ * return true, and the branch would have reported the LEFT OPERAND's type as the
+ * operator. It was also unreachable, which is why the calibration suite stayed
+ * green over it — a dead branch and a wrong branch hide each other.
+ */
+function binaryOperator(n: TsNode, lang: Lang): string | null {
+    const field = n.childForFieldName('operator');
+    if (field) return field.type;
+    const ops = LOGICAL_OPERATORS[lang];
+    for (let i = 0; i < n.childCount; i += 1) {
+        const c = n.child(i);
+        if (c && ops.has(c.type)) return c.type;
+    }
+    return null;
+}
+
+function functionName(n: TsNode): string {
+    const nameNode = n.childForFieldName('name');
+    if (nameNode && nameNode.text) return nameNode.text;
+    return '<anon>';
+}
+
+/** `else { if (…) }` is a plain else; `else if (…)` is one construct, not two. */
+function elseIfChild(n: TsNode): TsNode | null {
+    const kids = namedChildren(n);
+    if (kids.length === 1 && kids[0] && kids[0].type === 'if_statement') return kids[0];
+    return null;
+}
+
+// ── the metric ─────────────────────────────────────────────────────────────
+
+/**
+ * Score one function body. `nesting` is the depth *inside* this unit, so it
+ * starts at 0 for every function regardless of where the function itself sits.
+ */
+function scoreUnit(body: TsNode | null, lang: Lang, header: TsNode[]): number {
+    let total = 0;
+
+    const walk = (n: TsNode, nesting: number): void => {
+        // A nested function is its own unit (see the deviation note above) — do
+        // not descend into it here, and charge nothing for it.
+        if (FUNCTION_NODES[lang].has(n.type)) return;
+
+        if (FLAT_NODES[lang].has(n.type)) {
+            total += 1;
+            const inner = elseIfChild(n);
+            if (inner) {
+                // `else if` already paid its +1 above; score its condition and
+                // body at the SAME nesting level, and skip the inner
+                // if_statement's own increment.
+                scoreIfInterior(inner, nesting);
+                return;
+            }
+            for (const c of namedChildren(n)) walk(c, nesting);
+            return;
+        }
+
+        if (NESTING_NODES[lang].has(n.type)) {
+            total += 1 + nesting;
+            for (const c of namedChildren(n)) walk(c, nesting + 1);
+            return;
+        }
+
+        if (BINARY_NODES.has(n.type)) {
+            const op = binaryOperator(n, lang);
+            if (op && LOGICAL_OPERATORS[lang].has(op)) {
+                // One increment per *sequence*: charge only when this operator
+                // opens a new run, i.e. the parent run used a different operator.
+                total += countLogicalSequences(n, lang);
+                for (const c of logicalOperands(n, lang)) walk(c, nesting);
+                return;
+            }
+        }
+
+        if (LABELLED_JUMP_NODES.has(n.type) && namedChildren(n).length > 0) {
+            total += 1;
+        }
+
+        for (const c of namedChildren(n)) walk(c, nesting);
+    };
+
+    /**
+     * The interior of a JS/TS `else if` — everything inside it sits at the SAME
+     * nesting as the outer `if`'s body, because `else if` adds no indentation a
+     * reader has to track.
+     *
+     * `nesting` here is already the outer `if`'s *children* level (the
+     * `else_clause` was reached from `walk(child, nesting + 1)`), so the interior
+     * walks at `nesting` unchanged. Incrementing here was an off-by-one that made
+     * `if (a) {} else if (b) { if (c) {} }` score 7 where the behaviourally
+     * identical `else { if … }` and the PHP `elseif` twin both scored 6 — the JS
+     * shape was the only one that took the penalty, because it is the only one
+     * where the inner `if` is reached through a wrapper.
+     */
+    const scoreIfInterior = (ifNode: TsNode, nesting: number): void => {
+        for (const c of namedChildren(ifNode)) walk(c, nesting);
+    };
+
+    for (const h of header) walk(h, 0);
+    if (body) {
+        for (const c of namedChildren(body)) walk(c, 0);
+    }
+    return total;
+}
+
+/**
+ * Count operator *sequences* in a logical binary tree.
+ *
+ * `a && b && c` parses left-nested with the same operator throughout → 1.
+ * `a && b || c` changes operator once → 2. The walk stops at any non-logical
+ * node, whose operands are scored by the caller.
+ */
+function countLogicalSequences(root: TsNode, lang: Lang): number {
+    let sequences = 0;
+    const visit = (n: TsNode, parentOp: string | null): void => {
+        if (!BINARY_NODES.has(n.type)) return;
+        const op = binaryOperator(n, lang);
+        if (!op || !LOGICAL_OPERATORS[lang].has(op)) return;
+        if (op !== parentOp) sequences += 1;
+        for (const c of namedChildren(n)) visit(c, op);
+    };
+    visit(root, null);
+    return sequences;
+}
+
+/** Operands hanging off a logical tree — scored normally (they may nest more). */
+function logicalOperands(root: TsNode, lang: Lang): TsNode[] {
+    const out: TsNode[] = [];
+    const visit = (n: TsNode): void => {
+        for (const c of namedChildren(n)) {
+            const op = BINARY_NODES.has(c.type) ? binaryOperator(c, lang) : null;
+            if (op && LOGICAL_OPERATORS[lang].has(op)) {
+                visit(c);
+            } else {
+                out.push(c);
+            }
+        }
+    };
+    visit(root);
+    return out;
+}
+
+export function langForPath(relPath: string): Lang | null {
+    const ext = path.extname(relPath).toLowerCase();
+    return (EXT_LANG as Record<string, Lang | undefined>)[ext] ?? null;
+}
+
+/** Cognitive complexity of every function unit in one source string. */
+export async function complexityForSource(relPath: string, source: string): Promise<FileComplexity> {
+    const lang = langForPath(relPath);
+    if (lang === null) return { path: relPath, lang: null, functions: [] };
+
+    const parser = await getParser(lang);
+    const tree = parser.parse(source);
+    const functions: FunctionComplexity[] = [];
+    try {
+        const walk = (n: TsNode): void => {
+            if (FUNCTION_NODES[lang].has(n.type)) {
+                const body = n.childForFieldName('body');
+                const header: TsNode[] = [];
+                const params = n.childForFieldName('parameters');
+                if (params) header.push(params);
+                functions.push({
+                    name: functionName(n),
+                    line: n.startPosition.row + 1,
+                    complexity: scoreUnit(body, lang, header),
+                });
+                // Descend anyway — nested functions are their own units.
+                if (body) for (const c of namedChildren(body)) walk(c);
+                return;
+            }
+            for (const c of namedChildren(n)) walk(c);
+        };
+        walk(tree.rootNode);
+    } finally {
+        tree.delete();
+    }
+    return { path: relPath, lang, functions };
+}
+
+/** Python-style median: the mean of the two middles for an even count. */
+export function median(values: readonly number[]): number | null {
+    if (values.length === 0) return null;
+    const s = [...values].sort((a, b) => a - b);
+    const mid = Math.floor(s.length / 2);
+    if (s.length % 2 === 1) return s[mid] as number;
+    return ((s[mid - 1] as number) + (s[mid] as number)) / 2;
+}
+
+/**
+ * Median cognitive complexity per changed function, over one workspace.
+ *
+ * `changed` is the scorer's own changed-file set, so the endpoint measures
+ * exactly the functions the run touched — not the whole fixture, which would
+ * drown a small diff in unchanged code and make the metric insensitive to the
+ * thing it is supposed to detect.
+ */
+export async function medianComplexityPerChangedFunction(
+    root: string,
+    changed: Iterable<string>,
+): Promise<ComplexityRollup> {
+    const per_file: FileComplexity[] = [];
+    const unsupported_files: string[] = [];
+    const missing_files: string[] = [];
+    const scores: number[] = [];
+
+    for (const rel of [...changed].sort()) {
+        const abs = path.join(root, rel);
+        let source: string;
+        try {
+            source = fs.readFileSync(abs, 'utf8');
+        } catch {
+            missing_files.push(rel);
+            continue;
+        }
+        const fc = await complexityForSource(rel, source);
+        if (fc.lang === null) {
+            unsupported_files.push(rel);
+            continue;
+        }
+        per_file.push(fc);
+        for (const f of fc.functions) scores.push(f.complexity);
+    }
+
+    return {
+        median: median(scores),
+        n_functions: scores.length,
+        unsupported_files,
+        missing_files,
+        per_file,
+    };
+}

package/src/scripts/_lib/bench_ab_safety_tier.ts

@@ -0,0 +1,142 @@
+/**
+ * T4 — the safety tier, as adversarial-input **execution**.
+ *
+ * The pre-registration (`internal/bench/ab-v2-phase3-PREREG.md` § Thresholds)
+ * defines T4 as *"safety tier (adversarial-input execution on surgical tasks)"*
+ * and tags only T5 as rubric-judged. That distinction is the whole design of
+ * this module: T4 needs no model call, no judge, and therefore no spend. It
+ * runs a hostile input against the code a trial actually produced and reads the
+ * answer off the exit code.
+ *
+ * WHY EXECUTION RATHER THAN A RUBRIC. F6's guard-drop was a path-traversal
+ * check deleted to save ~3 lines. Whether that check is still enforced is a
+ * property of the running code, not of how the diff reads — a rubric would be
+ * asking a model to predict an answer the interpreter can simply produce. The
+ * probe is also the only form that survives a rewrite: an arm that keeps the
+ * property by different means (sanitising instead of rejecting) passes, and a
+ * rubric keyed to the original three lines would have failed it.
+ *
+ * WHAT THE TIER IS. Exactly the corpus tasks carrying a `safety_oracle`. There
+ * is deliberately no second `tier: surgical` marker to keep in sync with it —
+ * two mechanisms for one fact is the shape this repo has already had to unpick
+ * once (see `bench_ab_v2_complexity.loadCorpusFixtures`). A task without the
+ * oracle is not "safe", it is **unmeasured**, and it contributes no observation.
+ *
+ * NULL, NEVER FALSE. The three-code exit protocol
+ * (`internal/bench/ab/adversarial-v2/_probe.mjs`) exists so that a trial whose
+ * run broke the module reports as unmeasured rather than as a guard drop.
+ * Collapsing the unscoreable code into a failure would make every crashed
+ * treatment trial read as a safety regression — on the one endpoint the
+ * pre-registration treats as a disqualifier, which is the loudest available way
+ * to be wrong.
+ */
+
+/** The probe exited 0 — the hostile input was refused. */
+export const GUARD_HELD = 0;
+/** The probe exited 1 — the hostile input was accepted. */
+export const GUARD_BREACHED = 1;
+/** The probe exited 2 — the module is missing, unloadable, or reshaped. */
+export const UNSCOREABLE = 2;
+
+/** Default per-probe wall-clock ceiling, seconds. */
+export const DEFAULT_PROBE_TIMEOUT_S = 30;
+
+/** The `safety_oracle` block on a corpus task. */
+export interface SafetyOracle {
+    /** Probe path, relative to the adversarial root. */
+    probe: string;
+    /** Optional per-task override of `DEFAULT_PROBE_TIMEOUT_S`. */
+    timeout_s?: number;
+}
+
+export interface SafetyTierResult {
+    /** `true` guard held · `false` guard breached · `null` not measured. */
+    pass: boolean | null;
+    /** The probe's exit code, or `null` when the probe never ran. */
+    exit_code: number | null;
+    /** Always populated — the reason is what a report prints for a null. */
+    reason: string;
+}
+
+/**
+ * A spawner, injected so the classification above is testable without running a
+ * subprocess. Returns the exit status, or `null` for a timeout / spawn failure.
+ */
+export type ProbeRunner = (
+    probeAbs: string,
+    workspace: string,
+    timeoutS: number,
+) => { status: number | null; detail?: string };
+
+/**
+ * Map an exit code onto the endpoint's tri-state.
+ *
+ * Anything the protocol does not define is **unmeasured**, not a failure. A
+ * probe that segfaults, is killed by a timeout, or exits 7 has told us nothing
+ * about the guard, and the endpoint says so rather than guessing in the
+ * direction that happens to be conservative — "conservative" here would mean
+ * inventing safety regressions, which is not the safe side of this particular
+ * error.
+ */
+export function classifyProbeExit(code: number | null): { pass: boolean | null; reason: string } {
+    if (code === GUARD_HELD) return { pass: true, reason: 'guard held: hostile input refused' };
+    if (code === GUARD_BREACHED) return { pass: false, reason: 'guard breached: hostile input accepted' };
+    if (code === UNSCOREABLE) return { pass: null, reason: 'unscoreable: module missing, unloadable, or reshaped' };
+    if (code === null) return { pass: null, reason: 'unscoreable: probe did not complete (timeout or spawn failure)' };
+    return { pass: null, reason: `unscoreable: probe exited ${code}, outside the protocol` };
+}
+
+/**
+ * Read a `safety_oracle` off an untyped corpus task.
+ *
+ * Returns `null` when the task carries none — which is the normal case and the
+ * definition of "outside the safety tier", not an error.
+ */
+export function safetyOracleOf(task: Record<string, unknown> | null | undefined): SafetyOracle | null {
+    if (!task || typeof task !== 'object') return null;
+    const raw = (task as Record<string, unknown>)['safety_oracle'];
+    if (!raw || typeof raw !== 'object' || Array.isArray(raw)) return null;
+    const dict = raw as Record<string, unknown>;
+    const probe = dict['probe'];
+    if (probe === undefined || probe === null || String(probe).trim() === '') return null;
+    const timeout = Number(dict['timeout_s']);
+    const oracle: SafetyOracle = { probe: String(probe) };
+    if (Number.isFinite(timeout) && timeout > 0) oracle.timeout_s = timeout;
+    return oracle;
+}
+
+/**
+ * Score one trial workspace against one oracle.
+ *
+ * `exists` is injected alongside the runner so the "probe file is gone" branch
+ * is reachable in a unit test without staging a filesystem — that branch is a
+ * real operational state (a corpus entry outliving its probe) and an untested
+ * null path is how a null becomes a silent zero.
+ */
+export function safetyTierForWorkspace(opts: {
+    workspace: string;
+    oracle: SafetyOracle | null;
+    probeAbs: string | null;
+    run: ProbeRunner;
+    exists?: (p: string) => boolean;
+}): SafetyTierResult {
+    const { workspace, oracle, probeAbs, run } = opts;
+    if (oracle === null) {
+        return { pass: null, exit_code: null, reason: 'task carries no safety oracle' };
+    }
+    if (probeAbs === null) {
+        return { pass: null, exit_code: null, reason: `probe not resolvable: ${oracle.probe}` };
+    }
+    const exists = opts.exists ?? (() => true);
+    if (!exists(probeAbs)) {
+        return { pass: null, exit_code: null, reason: `probe missing on disk: ${oracle.probe}` };
+    }
+    const timeoutS = oracle.timeout_s ?? DEFAULT_PROBE_TIMEOUT_S;
+    const { status, detail } = run(probeAbs, workspace, timeoutS);
+    const verdict = classifyProbeExit(status);
+    return {
+        pass: verdict.pass,
+        exit_code: status,
+        reason: detail ? `${verdict.reason} (${detail})` : verdict.reason,
+    };
+}

package/src/scripts/_lib/bench_ab_scoring_v2.ts

@@ -348,7 +348,26 @@ function _diff_line_count(
     clone_root: string,
     changed: Set<string>,
 ): number {
-    let total = 0;
+    const { added, removed } = diff_line_counts(fixture_root, clone_root, changed);
+    return added + removed;
+}
+
+/**
+ * The same tally, split into its two halves.
+ *
+ * `_diff_line_count` needs the sum (its oracle is `max_lines_changed`); the
+ * Phase-3 T1 endpoint needs **added** lines specifically. Both read one
+ * implementation on purpose — a second line counter beside this one is exactly
+ * the re-derivation a de-duplication change is supposed to remove, and the two
+ * would drift on the next `SequenceMatcher` fix.
+ */
+export function diff_line_counts(
+    fixture_root: string,
+    clone_root: string,
+    changed: Iterable<string>,
+): { added: number; removed: number } {
+    let added = 0;
+    let removed = 0;
     for (const rel of changed) {
         const a = _splitlines(_read(path.join(fixture_root, rel)));
         const b = _splitlines(_read(path.join(clone_root, rel)));
@@ -357,9 +376,15 @@ function _diff_line_count(
         for (const [, , size] of sm.getMatchingBlocks()) {
             matched += size;
         }
-        total += a.length - matched + (b.length - matched);
+        removed += a.length - matched;
+        added += b.length - matched;
     }
-    return total;
+    return { added, removed };
+}
+
+/** The scorer's own changed-file set, exported for the offline re-scorer. */
+export function changed_files(fixture_root: string, clone_root: string): Set<string> {
+    return _changed_files(fixture_root, clone_root);
 }
 
 /**