kibi-cli 2.4.0 → 2.5.0

package/dist/search-answer.js

@@ -0,0 +1,846 @@
+import { escapeAtom, normalizeEntityId, parseEntityFromList, parseListOfLists, parseTriples, } from "./prolog/codec.js";
+import { entityRowsGoal } from "./public/operations/discovery-entities.js";
+import { runOperationJsonQuery } from "./public/operations/prolog-json.js";
+import { loadMarkdownBody } from "./search-ranking.js";
+const CURRENT_REQ_STATUSES = new Set([
+    "open",
+    "in_progress",
+    "closed",
+    "active",
+    "approved",
+]);
+const SUPERSEDED_ADR_STATUSES = new Set([
+    "superseded",
+    "deprecated",
+    "rejected",
+]);
+const ANSWER_RELATIONSHIPS = [
+    "specified_by",
+    "verified_by",
+    "validates",
+    "implements",
+    "constrains",
+    "requires_property",
+    "requires_predicate",
+    "requires_rule",
+    "supersedes",
+    "relates_to",
+    "exempts",
+];
+const FACT_LINKS = new Set([
+    "constrains",
+    "requires_property",
+    "requires_predicate",
+    "requires_rule",
+]);
+/** Properties the answer reads from a linked entity. */
+const ROW_KEYS = [
+    "id",
+    "title",
+    "status",
+    "fact_kind",
+    "approved_by",
+    "source",
+    "sourceFile",
+];
+// implements REQ-kibi-search-answer-layer-v2
+export const SEARCH_ANSWER_LIMITS = {
+    seeds: 10,
+    governing: 5,
+    perRequirement: 4,
+    rationale: 3,
+    notGoverning: 5,
+    observations: 3,
+    /** Entity rows loaded per answer, the matched seeds included. */
+    entityLoads: 70,
+    maxBytes: 16_384,
+    /** Title length (characters) kept when an answer must shrink to fit. */
+    titleChars: 160,
+    /** Characters of an ADR decision kept as its excerpt. */
+    excerptChars: 280,
+    /** Characters of a witness or unknown detail kept. */
+    detailChars: 240,
+};
+function text(value) {
+    return typeof value === "string" ? value : value == null ? "" : String(value);
+}
+function unique(ids) {
+    return [...new Set(ids.filter((id) => id.length > 0))];
+}
+function brief(entity) {
+    const status = text(entity.status);
+    return {
+        id: text(entity.id),
+        title: text(entity.title),
+        ...(status ? { status } : {}),
+    };
+}
+function edgesGoal(ids) {
+    const idList = ids.map((id) => `'${escapeAtom(id)}'`).join(",");
+    return `findall([Rel,From,To], (member(Rel, [${ANSWER_RELATIONSHIPS.join(",")}]), kb_relationship(Rel, From, To), (member(From, [${idList}]); member(To, [${idList}]))), Edges)`;
+}
+/**
+ * The part of the KB graph the answer reads, fetched in batches: each
+ * `fetch` is one engine round trip for every id whose edges or row are not
+ * known yet. Answering used to cost one round trip per edge hop and per
+ * linked entity; on a journaled engine each costs a polling interval, which
+ * made a warm search take seconds.
+ */
+class AnswerGraph {
+    prolog;
+    rowBudget;
+    edgesById = new Map();
+    rows = new Map();
+    rowLoads = 0;
+    truncated = false;
+    constructor(prolog, rowBudget) {
+        this.prolog = prolog;
+        this.rowBudget = rowBudget;
+    }
+    seed(entity) {
+        const id = text(entity.id);
+        if (id)
+            this.rows.set(id, { ...entity });
+    }
+    row(id) {
+        return this.rows.get(id) ?? null;
+    }
+    /** Edges touching any of `ids` that were fetched, without duplicates. */
+    edgesOf(ids) {
+        const seen = new Set();
+        const edges = [];
+        for (const id of ids) {
+            for (const edge of this.edgesById.get(id) ?? []) {
+                const key = `${edge.rel}|${edge.from}|${edge.to}`;
+                if (seen.has(key))
+                    continue;
+                seen.add(key);
+                edges.push(edge);
+            }
+        }
+        return edges;
+    }
+    /**
+     * Fetch missing edges and rows in one query. `alsoGoal` is conjoined so a
+     * further lookup shares the round trip; its bindings are returned.
+     */
+    async fetch(request) {
+        const edgeIds = unique(request.edges ?? []).filter((id) => !this.edgesById.has(id));
+        const wanted = unique(request.rows ?? []).filter((id) => !this.rows.has(id));
+        const room = Math.max(0, this.rowBudget - this.rowLoads);
+        if (wanted.length > room)
+            this.truncated = true;
+        const rowIds = wanted.slice(0, room);
+        const goals = [
+            ...(edgeIds.length > 0 ? [edgesGoal(edgeIds)] : []),
+            ...(rowIds.length > 0 ? [entityRowsGoal(rowIds, ROW_KEYS)] : []),
+            ...(request.alsoGoal ? [request.alsoGoal] : []),
+        ];
+        if (goals.length === 0)
+            return {};
+        const result = await this.prolog.query(goals.join(", "));
+        if (!result.success) {
+            // A failed edge lookup reads as no links, as before batching; a
+            // failed entity lookup must not silently drop requirements.
+            if (rowIds.length > 0 || request.alsoGoal) {
+                throw new Error(result.error || "Search answer graph query failed");
+            }
+            for (const id of edgeIds)
+                this.edgesById.set(id, []);
+            return {};
+        }
+        if (edgeIds.length > 0) {
+            for (const id of edgeIds)
+                this.edgesById.set(id, []);
+            for (const [rel, from, to] of parseTriples(result.bindings.Edges ?? "[]")) {
+                const edge = {
+                    rel,
+                    from: normalizeEntityId(from),
+                    to: normalizeEntityId(to),
+                };
+                this.edgesById.get(edge.from)?.push(edge);
+                if (edge.to !== edge.from)
+                    this.edgesById.get(edge.to)?.push(edge);
+            }
+        }
+        if (rowIds.length > 0) {
+            this.rowLoads += rowIds.length;
+            for (const id of rowIds)
+                this.rows.set(id, null);
+            for (const row of parseListOfLists(result.bindings.Results ?? "[]").map(parseEntityFromList))
+                this.rows.set(normalizeEntityId(text(row.id)), row);
+        }
+        return result.bindings;
+    }
+}
+/** Tests that verify a scenario: `verified_by` from it or `validates` into it. */
+function scenarioTests(scenarioId, edges) {
+    const tests = [];
+    for (const edge of edges) {
+        if (edge.rel === "verified_by" && edge.from === scenarioId)
+            tests.push(edge.to);
+        if (edge.rel === "validates" && edge.to === scenarioId)
+            tests.push(edge.from);
+    }
+    return tests;
+}
+/** Requirement ids a non-requirement entity points back to. */
+function owningRequirements(id, type, edges) {
+    const owners = [];
+    for (const edge of edges) {
+        if (type === "scenario" && edge.rel === "specified_by" && edge.to === id)
+            owners.push(edge.from);
+        if (type === "test" && edge.rel === "verified_by" && edge.to === id)
+            owners.push(edge.from);
+        if (type === "test" && edge.rel === "validates" && edge.from === id)
+            owners.push(edge.to);
+        if (type === "symbol" && edge.rel === "implements" && edge.from === id)
+            owners.push(edge.to);
+        if (type === "fact" && FACT_LINKS.has(edge.rel) && edge.to === id)
+            owners.push(edge.from);
+    }
+    return owners;
+}
+/** Shorten to at most `maxChars` code points, ending with an ellipsis. */
+function clip(value, maxChars) {
+    const chars = Array.from(value);
+    if (chars.length <= maxChars)
+        return value;
+    return `${chars.slice(0, Math.max(0, maxChars - 1)).join("")}…`;
+}
+/**
+ * The opening sentences of an ADR's Decision section (falling back to
+ * Rationale, then to the first prose paragraph), as plain text of at most
+ * `maxChars` characters.
+ */
+// implements REQ-kibi-search-answer-layer-v2
+export function adrExcerpt(body, maxChars = SEARCH_ANSWER_LIMITS.excerptChars) {
+    const sections = new Map();
+    const preamble = [];
+    let current = preamble;
+    let fenced = false;
+    for (const line of body.split(/\r?\n/)) {
+        if (/^\s*(```|~~~)/.test(line)) {
+            fenced = !fenced;
+            continue;
+        }
+        if (fenced)
+            continue;
+        const heading = /^\s{0,3}#{1,6}\s+(.*?)\s*#*\s*$/.exec(line);
+        if (heading) {
+            const name = text(heading[1]).toLowerCase();
+            current = [];
+            if (!sections.has(name))
+                sections.set(name, current);
+            continue;
+        }
+        current.push(line);
+    }
+    const named = (pattern) => [...sections.entries()].find(([name, lines]) => pattern.test(name) && lines.join("").trim() !== "")?.[1];
+    const lines = named(/^decision\b/) ??
+        named(/^rationale\b/) ??
+        (preamble.join("").trim() !== ""
+            ? preamble
+            : [...sections.values()].find((section) => section.join("").trim()));
+    if (!lines)
+        return undefined;
+    const prose = lines
+        .map((line) => line
+        .replace(/^\s*(?:[-*+]|\d+[.)])\s+/, "")
+        .replace(/^\s*>\s?/, "")
+        .replace(/!?\[([^\]]*)\]\([^)]*\)/g, "$1")
+        .replace(/\*\*|__|`/g, "")
+        .trim())
+        .filter((line) => line !== "" && !/^\|/.test(line))
+        .join(" ")
+        .replace(/\s+/g, " ")
+        .trim();
+    if (!prose)
+        return undefined;
+    const sentences = prose.split(/(?<=[.!?])\s+/);
+    let excerpt = "";
+    for (const sentence of sentences) {
+        const next = `${excerpt} ${sentence.trim()}`.trim();
+        if (excerpt !== "" && Array.from(next).length > maxChars)
+            break;
+        excerpt = next;
+    }
+    return clip(excerpt, maxChars);
+}
+async function rationaleOf(row, workspaceRoot, bodies) {
+    const source = text(row.sourceFile) || text(row.source);
+    if (!source)
+        return brief(row);
+    let excerpt;
+    if (workspaceRoot !== undefined) {
+        let body = bodies.get(source);
+        if (body === undefined) {
+            body = loadMarkdownBody(source, workspaceRoot);
+            bodies.set(source, body);
+        }
+        const markdown = await body;
+        excerpt = markdown ? adrExcerpt(markdown) : undefined;
+    }
+    return {
+        ...brief(row),
+        source: source.split("#", 1)[0] ?? source,
+        ...(excerpt ? { excerpt } : {}),
+    };
+}
+function records(value) {
+    return Array.isArray(value)
+        ? value.filter((item) => item !== null && typeof item === "object" && !Array.isArray(item))
+        : [];
+}
+function strings(value) {
+    return Array.isArray(value)
+        ? value.map((item) => normalizeEntityId(text(item))).filter(Boolean)
+        : [];
+}
+const FEASIBILITY_UNKNOWN_TEXT = {
+    no_assumptions: "expects success but assumes nothing, so its feasibility against current requirements is unknown",
+    unmatched_assumption: "expects success but an assumption is not constrained by any current requirement, so its feasibility is unknown",
+    incomparable_assumption: "expects success but an assumption cannot be compared with the requirement constraint (type or unit differs), so its feasibility is unknown",
+    contradictory_assumptions: "expects success but its assumptions contradict each other, so the scenario can never happen",
+    conflicting_requirements: "expects success but the requirements constraining it conflict with each other, so its feasibility is unknown",
+    undecided_rule: "expects success but a rule governing it can be neither satisfied nor refuted from its assumptions, so its feasibility is unknown",
+    undetermined_validity: "expects success but would conflict only with constraints whose validity window may not cover the scenario's time, so its feasibility is unknown",
+};
+/** Project the Prolog verdict of one requirement onto the answer fields. */
+// implements REQ-kibi-search-answer-layer-v2
+export function projectVerdict(raw, perRequirement = SEARCH_ANSWER_LIMITS.perRequirement) {
+    const detail = (value) => clip(text(value), SEARCH_ANSWER_LIMITS.detailChars);
+    const witnesses = [];
+    const unknowns = [];
+    if (raw === undefined || raw.error !== undefined) {
+        unknowns.push({
+            kind: "verdict_unavailable",
+            detail: detail(raw?.error !== undefined
+                ? `The checks could not be evaluated for this requirement: ${text(raw.error)}`
+                : "The checks could not be evaluated for this requirement."),
+            entities: [],
+        });
+    }
+    else {
+        for (const row of records(raw.contradictions)) {
+            const other = normalizeEntityId(text(row.with));
+            if (text(row.status) === "contradiction") {
+                witnesses.push({
+                    check: "domain-contradictions",
+                    status: "contradiction",
+                    ...(other ? { with: other } : {}),
+                    facts: strings(row.facts),
+                    detail: detail(row.reason),
+                });
+            }
+            else {
+                unknowns.push({
+                    kind: "contradiction_unresolved",
+                    detail: detail(row.reason),
+                    entities: [other, ...strings(row.facts)].filter(Boolean),
+                });
+            }
+        }
+        for (const row of records(raw.scenarios)) {
+            const scenario = normalizeEntityId(text(row.scenario));
+            const outcome = text(row.outcome);
+            if (outcome === "infeasible") {
+                const forbidding = normalizeEntityId(text(row.requirement));
+                witnesses.push({
+                    check: "scenario-feasibility",
+                    status: "infeasible",
+                    ...(forbidding ? { with: forbidding } : {}),
+                    scenario,
+                    facts: [...strings(row.assumed), ...strings(row.fact)],
+                    detail: detail(row.reason),
+                });
+            }
+            else if (outcome === "unknown") {
+                const reason = text(row.reason);
+                unknowns.push({
+                    kind: "feasibility_unknown",
+                    detail: detail(`${scenario} ${FEASIBILITY_UNKNOWN_TEXT[reason] ?? `has unknown feasibility (${reason})`}`),
+                    entities: [scenario, ...strings(row.facts)],
+                });
+            }
+        }
+        for (const row of records(raw.forbids)) {
+            witnesses.push({
+                check: "scenario-feasibility",
+                status: "infeasible",
+                scenario: normalizeEntityId(text(row.scenario)),
+                facts: [...strings(row.assumed), ...strings(row.fact)],
+                detail: detail(row.reason),
+            });
+        }
+        const inventory = records([raw.inventory])[0];
+        if (inventory !== undefined) {
+            for (const claim of records(inventory.unresolved)) {
+                unknowns.push({
+                    kind: "unresolved_proposition",
+                    detail: detail(`${text(claim.status)}: ${text(claim.claim)}`),
+                    entities: [],
+                });
+            }
+            if (text(inventory.status) === "missing") {
+                unknowns.push({
+                    kind: "analysis_incomplete",
+                    detail: "The requirement's clause ledger (semantic inventory) is missing or incomplete, so contradiction analysis cannot cover all of its clauses.",
+                    entities: [],
+                });
+            }
+        }
+    }
+    const status = witnesses.some((witness) => witness.status === "contradiction")
+        ? "contradiction"
+        : witnesses.length > 0
+            ? "infeasible"
+            : unknowns.length > 0
+                ? "unknown"
+                : "none";
+    // Contradictions first: they block regardless of scenarios.
+    witnesses.sort((left, right) => left.status === right.status ? 0 : left.status === "contradiction" ? -1 : 1);
+    return {
+        verdict: { status, witnesses: witnesses.slice(0, perRequirement) },
+        unknowns: unknowns.slice(0, perRequirement),
+        truncated: witnesses.length > perRequirement || unknowns.length > perRequirement,
+    };
+}
+const UNKNOWN_PROCEDURE = /existence_error|[Uu]nknown procedure|does not exist/;
+function decodeJsonBinding(raw) {
+    if (raw === undefined)
+        return undefined;
+    let parsed = JSON.parse(raw);
+    if (typeof parsed === "string")
+        parsed = JSON.parse(parsed);
+    return parsed;
+}
+/**
+ * Rows the answer still needs plus the verdicts of the governing
+ * requirements, sharing one round trip while the discovery module is loaded;
+ * otherwise the module is loaded first. A verdict that cannot be computed
+ * leaves the answer standing with `verdict_unavailable` unknowns.
+ */
+async function fetchVerdicts(prolog, graph, rowIds, reqIds) {
+    const goal = `discovery:search_answer_verdicts_json([${reqIds.map((id) => `'${escapeAtom(id)}'`).join(",")}], JsonString)`;
+    try {
+        const bindings = await graph.fetch({ rows: rowIds, alsoGoal: goal });
+        return decodeJsonBinding(bindings.JsonString);
+    }
+    catch (error) {
+        const message = error instanceof Error ? error.message : String(error);
+        await graph.fetch({ rows: rowIds });
+        if (!UNKNOWN_PROCEDURE.test(message))
+            return undefined;
+    }
+    try {
+        return await runOperationJsonQuery(prolog, "discovery.pl", goal, "Search answer verdicts");
+    }
+    catch {
+        return undefined;
+    }
+}
+export async function buildSearchAnswer(prolog, matches, options = {}) {
+    const limits = SEARCH_ANSWER_LIMITS;
+    const graph = new AnswerGraph(prolog, limits.entityLoads);
+    const seeds = matches.slice(0, limits.seeds);
+    let truncated = matches.length > limits.seeds;
+    const seedIds = seeds.map((match) => text(match.entity.id));
+    // Search rows are projected without fact_kind or approved_by, so the
+    // matched entities' own rows are loaded with their edges: notes are told
+    // apart from rules, and a matched exception keeps its approver. A match
+    // the store no longer holds keeps its search row.
+    await graph.fetch({ edges: seedIds, rows: seedIds });
+    for (const match of seeds)
+        if (graph.row(text(match.entity.id)) === null)
+            graph.seed(match.entity);
+    const seedEdges = graph.edgesOf(seedIds);
+    // Candidate requirements: matched requirements first, then requirements
+    // that own a matched scenario, test, symbol or fact.
+    const candidates = new Map();
+    const propose = (owner, score, via) => {
+        const previous = candidates.get(owner);
+        if (!previous || previous.score < score)
+            candidates.set(owner, { score, via });
+    };
+    const observations = [];
+    const rationaleRows = new Map();
+    // A matched test may verify a scenario rather than a requirement; its
+    // targets are resolved to their governing requirements after the loop.
+    const testTargets = [];
+    for (const match of seeds) {
+        const id = text(match.entity.id);
+        const entity = graph.row(id) ?? match.entity;
+        const type = text(entity.type || match.entity.type);
+        if (type === "req") {
+            if (!candidates.has(id))
+                candidates.set(id, { score: match.score, via: "matched" });
+            continue;
+        }
+        if (type === "adr") {
+            if (!SUPERSEDED_ADR_STATUSES.has(text(entity.status)))
+                rationaleRows.set(id, entity);
+            continue;
+        }
+        if (type === "fact") {
+            const kind = text(entity.fact_kind);
+            if (kind === "observation" || kind === "meta") {
+                observations.push(brief(entity));
+                continue;
+            }
+        }
+        const owners = owningRequirements(id, type, seedEdges);
+        if (type === "test") {
+            testTargets.push({ id, score: match.score, targets: owners });
+            continue;
+        }
+        for (const owner of owners)
+            propose(owner, match.score * 0.9, `${type} ${id}`);
+    }
+    const byScore = () => [...candidates.entries()]
+        .sort(([, left], [, right]) => right.score - left.score)
+        .map(([id]) => id);
+    const targetIds = unique(testTargets.flatMap((test) => test.targets));
+    // Test targets and the candidates found so far share one round trip.
+    await graph.fetch({
+        edges: [...targetIds, ...byScore()],
+        rows: byScore(),
+    });
+    const targetEdges = graph.edgesOf(targetIds);
+    for (const test of testTargets) {
+        for (const target of test.targets) {
+            // A target that some requirement specifies is a scenario: lift the
+            // test to that requirement. Anything else is a directly verified
+            // requirement (non-requirements are filtered when loaded).
+            const scenarioOwners = targetEdges
+                .filter((edge) => edge.rel === "specified_by" && edge.to === target)
+                .map((edge) => edge.from);
+            if (scenarioOwners.length === 0) {
+                propose(target, test.score * 0.9, `test ${test.id}`);
+                continue;
+            }
+            for (const owner of scenarioOwners)
+                propose(owner, test.score * 0.85, `test ${test.id} via ${target}`);
+        }
+    }
+    const supersededBy = new Map();
+    for (const edge of seedEdges)
+        if (edge.rel === "supersedes")
+            supersededBy.set(edge.to, edge.from);
+    // A matched requirement that was superseded points at what governs now:
+    // follow the chain so the replacement is a candidate too.
+    let frontier = byScore();
+    for (let hop = 0; hop < 3 && frontier.length > 0; hop += 1) {
+        await graph.fetch({ edges: frontier, rows: frontier });
+        for (const edge of graph.edgesOf(frontier))
+            if (edge.rel === "supersedes")
+                supersededBy.set(edge.to, edge.from);
+        const next = [];
+        for (const id of frontier) {
+            const replacement = supersededBy.get(id);
+            const source = candidates.get(id);
+            if (replacement === undefined || source === undefined)
+                continue;
+            if (!candidates.has(replacement)) {
+                candidates.set(replacement, {
+                    score: source.score * 0.95,
+                    via: `supersedes ${id}`,
+                });
+                next.push(replacement);
+            }
+        }
+        frontier = next;
+    }
+    // A replacement found on the last hop still needs its row and links, and
+    // a replacement that was itself superseded must not read as current.
+    if (frontier.length > 0) {
+        await graph.fetch({ edges: frontier, rows: frontier });
+        for (const edge of graph.edgesOf(frontier))
+            if (edge.rel === "supersedes")
+                supersededBy.set(edge.to, edge.from);
+    }
+    const chosen = [];
+    const notGoverning = [];
+    const ordered = [...candidates.entries()].sort(([leftId, left], [rightId, right]) => right.score - left.score || leftId.localeCompare(rightId));
+    for (const [id, candidate] of ordered) {
+        const entity = graph.row(id);
+        if (!entity || text(entity.type || "req") !== "req")
+            continue;
+        const replacement = supersededBy.get(id);
+        if (replacement !== undefined ||
+            !CURRENT_REQ_STATUSES.has(text(entity.status))) {
+            if (notGoverning.length < limits.notGoverning)
+                notGoverning.push({
+                    ...brief(entity),
+                    ...(replacement !== undefined ? { supersededBy: replacement } : {}),
+                });
+            else
+                truncated = true;
+            continue;
+        }
+        if (chosen.length >= limits.governing) {
+            truncated = true;
+            continue;
+        }
+        chosen.push({ id, entity, ...candidate });
+    }
+    // What each governing requirement links to, from the edges fetched above.
+    const firstOf = (ids) => {
+        const distinct = unique(ids);
+        if (distinct.length > limits.perRequirement)
+            truncated = true;
+        return distinct.slice(0, limits.perRequirement);
+    };
+    const plans = chosen.map((req) => {
+        const edges = graph.edgesOf([req.id]);
+        const from = (rel) => edges
+            .filter((edge) => edge.rel === rel && edge.from === req.id)
+            .map((edge) => edge.to);
+        const into = (rel) => edges
+            .filter((edge) => edge.rel === rel && edge.to === req.id)
+            .map((edge) => edge.from);
+        const scenarioIds = unique(from("specified_by"));
+        const traversed = firstOf(scenarioIds);
+        return {
+            req,
+            facts: firstOf(edges
+                .filter((edge) => FACT_LINKS.has(edge.rel) && edge.from === req.id)
+                .map((edge) => edge.to)),
+            scenarios: traversed,
+            directTests: unique([...from("verified_by"), ...into("validates")]),
+            related: firstOf([...from("relates_to"), ...into("relates_to")]),
+            exceptions: firstOf(into("exempts")),
+        };
+    });
+    await graph.fetch({
+        edges: plans.flatMap((plan) => plan.scenarios),
+        rows: plans.flatMap((plan) => [
+            ...plan.facts,
+            ...plan.scenarios,
+            ...plan.related,
+            ...plan.exceptions,
+            ...plan.directTests.slice(0, limits.perRequirement),
+        ]),
+    });
+    // Tests linked to the requirement itself come first, then tests that
+    // verify one of its scenarios; a test reached both ways counts as direct.
+    const testVias = plans.map((plan) => {
+        const testVia = new Map();
+        for (const testId of plan.directTests)
+            if (!testVia.has(testId))
+                testVia.set(testId, "direct");
+        const verification = graph.edgesOf(plan.scenarios);
+        for (const scenarioId of plan.scenarios)
+            for (const testId of scenarioTests(scenarioId, verification))
+                if (!testVia.has(testId))
+                    testVia.set(testId, scenarioId);
+        return testVia;
+    });
+    const testIds = testVias.map((testVia) => firstOf([...testVia.keys()]));
+    const verdicts = await fetchVerdicts(prolog, graph, testIds.flat(), chosen.map((req) => req.id));
+    const verdictById = new Map(records(verdicts?.requirements).map((row) => [
+        normalizeEntityId(text(row.id)),
+        row,
+    ]));
+    const bodies = new Map();
+    const rowsOf = (ids) => ids.map((id) => graph.row(id)).filter((row) => row !== null);
+    const governing = [];
+    for (const [index, plan] of plans.entries()) {
+        const { req } = plan;
+        const adrRows = rowsOf(plan.related).filter((row) => text(row.type) === "adr" &&
+            !SUPERSEDED_ADR_STATUSES.has(text(row.status)));
+        for (const adr of adrRows)
+            rationaleRows.set(text(adr.id), adr);
+        const projected = projectVerdict(verdicts === undefined ? undefined : verdictById.get(req.id));
+        if (projected.truncated)
+            truncated = true;
+        const testVia = testVias[index] ?? new Map();
+        governing.push({
+            ...brief(req.entity),
+            score: req.score,
+            via: req.via,
+            facts: rowsOf(plan.facts).map((fact) => {
+                const factKind = text(fact.fact_kind);
+                return { ...brief(fact), ...(factKind ? { factKind } : {}) };
+            }),
+            scenarios: rowsOf(plan.scenarios).map(brief),
+            tests: rowsOf(testIds[index] ?? []).map((row) => ({
+                ...brief(row),
+                via: testVia.get(text(row.id)) ?? "direct",
+            })),
+            adrs: adrRows.map(brief),
+            verdict: projected.verdict,
+            exceptions: rowsOf(plan.exceptions).map((row) => {
+                const approvedBy = text(row.approved_by).trim();
+                return { ...brief(row), ...(approvedBy ? { approvedBy } : {}) };
+            }),
+            unknowns: projected.unknowns,
+        });
+    }
+    const rationaleList = [...rationaleRows.values()];
+    if (rationaleList.length > limits.rationale)
+        truncated = true;
+    if (observations.length > limits.observations)
+        truncated = true;
+    const rationale = await Promise.all(rationaleList
+        .slice(0, limits.rationale)
+        .map((row) => rationaleOf(row, options.workspaceRoot, bodies)));
+    const scope = {
+        branch: options.branch ?? null,
+        snapshotId: text(verdicts?.scope?.snapshotId) || "unknown",
+        syncedAt: typeof verdicts?.scope?.syncedAt === "string"
+            ? verdicts.scope.syncedAt
+            : null,
+    };
+    const answer = {
+        version: "kibi.search-answer.v1",
+        governing,
+        rationale,
+        notGoverning,
+        observations: observations.slice(0, limits.observations),
+        scope,
+        truncated: truncated || graph.truncated,
+        note: governing.length === 0
+            ? "No current requirement matched or owns a matched entity. Absence here is not evidence that nothing governs the change; refine the query or pass sourceLocations."
+            : "Governing requirements are current requirements matched by the query or linked to a matched entity. Links are discovery, not proof. A verdict lists the contradiction and scenario-feasibility witnesses that name the requirement; none means no check named it, not that it is proven. Use kb_check and kb_coverage for consistency and proof status.",
+    };
+    return fitSearchAnswer(answer, limits.maxBytes);
+}
+function byteSize(value) {
+    return Buffer.byteLength(JSON.stringify(value), "utf8");
+}
+function clipTitle(entity, chars) {
+    return { ...entity, title: clip(entity.title, chars) };
+}
+function clipRationale(entity, chars) {
+    const clipped = clipTitle(entity, chars);
+    return entity.excerpt === undefined
+        ? clipped
+        : {
+            ...clipped,
+            excerpt: clip(entity.excerpt, SEARCH_ANSWER_LIMITS.excerptChars),
+        };
+}
+function withoutExcerpt(entity) {
+    if (entity.excerpt === undefined)
+        return entity;
+    const { excerpt: _excerpt, ...rest } = entity;
+    return rest;
+}
+const ANSWER_TRIM_ORDER = [
+    "observations",
+    "notGoverning",
+    "rationale",
+];
+const REQUIREMENT_TRIM_ORDER = [
+    "unknowns",
+    "exceptions",
+    "adrs",
+    "tests",
+    "scenarios",
+    "facts",
+];
+/**
+ * Hold the whole answer under `maxBytes`, marking it truncated when anything
+ * is cut. In order: long titles, excerpts and details are clipped; ADR
+ * excerpts are dropped, last rationale entry first; the least relevant
+ * governing requirements are dropped;
+ * list items are dropped (notes and side lists before the remaining
+ * requirement's own lists, its verdict witnesses last); then the remaining
+ * requirement's title and `via` are clipped. A verdict keeps its status when
+ * its witnesses are dropped. If even that cannot fit, the requirement itself
+ * is dropped.
+ */
+export function fitSearchAnswer(answer, maxBytes = SEARCH_ANSWER_LIMITS.maxBytes) {
+    let current = answer;
+    const over = () => byteSize(current) > maxBytes;
+    if (!over())
+        return current;
+    const titleChars = SEARCH_ANSWER_LIMITS.titleChars;
+    const detailChars = SEARCH_ANSWER_LIMITS.detailChars;
+    const clipAll = (rows) => rows.map((row) => clipTitle(row, titleChars));
+    current = {
+        ...current,
+        truncated: true,
+        governing: current.governing.map((req) => ({
+            ...clipTitle(req, titleChars),
+            facts: clipAll(req.facts),
+            scenarios: clipAll(req.scenarios),
+            tests: clipAll(req.tests),
+            adrs: clipAll(req.adrs),
+            verdict: {
+                ...req.verdict,
+                witnesses: req.verdict.witnesses.map((witness) => ({
+                    ...witness,
+                    detail: clip(witness.detail, detailChars),
+                })),
+            },
+            exceptions: clipAll(req.exceptions),
+            unknowns: req.unknowns.map((unknown) => ({
+                ...unknown,
+                detail: clip(unknown.detail, detailChars),
+            })),
+        })),
+        rationale: current.rationale.map((adr) => clipRationale(adr, titleChars)),
+        notGoverning: clipAll(current.notGoverning),
+        observations: clipAll(current.observations),
+    };
+    // Excerpts are the most expendable text: drop them before any entry.
+    for (let index = current.rationale.length - 1; index >= 0 && over(); index -= 1) {
+        const rationale = [...current.rationale];
+        rationale[index] = withoutExcerpt(rationale[index]);
+        current = { ...current, rationale };
+    }
+    while (over() && current.governing.length > 1)
+        current = { ...current, governing: current.governing.slice(0, -1) };
+    while (over()) {
+        const side = ANSWER_TRIM_ORDER.find((key) => current[key].length > 0);
+        if (side) {
+            current = { ...current, [side]: current[side].slice(0, -1) };
+            continue;
+        }
+        const [req] = current.governing;
+        if (!req)
+            break;
+        const own = REQUIREMENT_TRIM_ORDER.find((key) => req[key].length > 0);
+        if (own) {
+            current = {
+                ...current,
+                governing: [{ ...req, [own]: req[own].slice(0, -1) }],
+            };
+            continue;
+        }
+        if (req.verdict.witnesses.length === 0)
+            break;
+        current = {
+            ...current,
+            governing: [
+                {
+                    ...req,
+                    verdict: {
+                        ...req.verdict,
+                        witnesses: req.verdict.witnesses.slice(0, -1),
+                    },
+                },
+            ],
+        };
+    }
+    for (const field of ["title", "via"]) {
+        while (over()) {
+            const [req] = current.governing;
+            const chars = req ? Array.from(req[field]).length : 0;
+            if (!req || chars <= 1)
+                break;
+            // Each removed code point frees at least one byte; one more makes room
+            // for the ellipsis.
+            const keep = Math.max(1, chars - (byteSize(current) - maxBytes) - 2);
+            current = {
+                ...current,
+                governing: [{ ...req, [field]: clip(req[field], keep) }],
+            };
+        }
+    }
+    if (over())
+        current = { ...current, governing: [] };
+    return current;
+}