@ecoma-io/archkeep 0.13.0 → 0.14.0

package/src/commands/delta-snapshot.mjs

@@ -0,0 +1,517 @@
+/**
+ * The evidence snapshot: what a future `delta <base.json>` run consumes as its
+ * baseline.
+ *
+ * The snapshot stores EVIDENCE, not verdicts. A baseline that stored "which
+ * violations existed at base" would be judged under the law and the clock of
+ * the moment it was captured, so a policy edit or a waiver expiring between
+ * base and head would fabricate classifications: a violation introduced by a
+ * policy tightening would read as "introduced by the code", and one whose
+ * waiver lapsed would read as "resolved" when it is live again. So the
+ * snapshot carries the raw import-site records (`../analysis/contract.md`),
+ * the graph they were collected against, and the coverage facts that say how
+ * complete the look was — everything `../rules/index.mjs`'s
+ * `evaluate(sites, graph, config)` needs to re-judge the base under the CURRENT
+ * config at delta time. Both sides are then judged under ONE law and ONE
+ * shared reference instant, and only the code can move a classification.
+ *
+ * The format has its own `schemaVersion`, independent of the report envelope's:
+ * this file is read back by `parseEvidenceSnapshot` alone, never by the report
+ * renderers, and the two formats will evolve on different clocks.
+ *
+ * ## Purity seam
+ *
+ * Everything decidable is pure: `buildEvidenceSnapshot` takes already-resolved
+ * records as arguments, `serializeEvidenceSnapshot` takes the snapshot object,
+ * and `parseEvidenceSnapshot` takes text. Only `readEvidenceSnapshot` touches
+ * the filesystem, and its read function is injectable — the same separation
+ * `./history.mjs` draws between `readSnapshots` and `computeEvolution`
+ * (`../../../../AGENTS.md`: gate logic takes its facts as arguments).
+ */
+import { readFileSync } from "node:fs";
+
+import { buildDependencies, buildProjects } from "./graph.mjs";
+
+/** The only snapshot schemaVersion this module writes and reads. */
+export const EVIDENCE_SNAPSHOT_SCHEMA_VERSION = 1;
+
+/**
+ * Builds the snapshot object from already-captured evidence.
+ *
+ * Every argument is a fact the caller already resolved — provenance from
+ * `./provenance.mjs`'s `resolveProvenance`, the fingerprint from
+ * `./graph.mjs`'s `computePolicyFingerprint`, the graph in Nx's shape, and the
+ * analysis envelope's `imports` array. Nothing here reads a file, spawns a
+ * process, or consults a clock, so a test drives capture without mocks.
+ *
+ * The graph is normalized through `buildProjects`/`buildDependencies` — the
+ * exact functions the `graph` command serializes with — so a snapshot's graph
+ * section cannot drift from what `graph --format json` publishes for the same
+ * tree. Three rule-relevant fields `buildProjects` strips for the public graph
+ * contract are re-attached here when the node declares them
+ * (`mfeRemote`, `entryPoints`, `declaredPackages`): `evaluate()` reads each,
+ * and a re-judge against a project row missing one would fabricate or silence
+ * verdicts — `declaredPackages` absent reads as "depends directly on nothing",
+ * which turns transitive-dependency violations on and off by omission. This
+ * snapshot exists to be re-judged from; it stores what judging reads.
+ *
+ * @param {object} input
+ * @param {{name: string, version: string}} input.tool The tool that captured
+ *   the snapshot, named so a reader can tell which engine produced the bytes.
+ * @param {{commit: string, remote: string|null, dirty: boolean}|null}
+ *   input.provenance From `resolveProvenance`; `null` is carried as an explicit
+ *   "no origin claim" rather than dropped.
+ * @param {string} input.provider The project-model provider that built the
+ *   graph ("nx", "native", "moon").
+ * @param {string} input.policyFingerprint From `computePolicyFingerprint`.
+ * @param {{complete: boolean, analyzedFiles: number, notAnalyzed: object[],
+ *   blindSpots: object[]}} input.coverage The coverage summary, partitioned
+ *   exactly as `./graph.mjs` partitions analysis failures.
+ * @param {{nodes: object, dependencies: object, workspaceLayout?: object,
+ *   exemptedFiles?: string[]}} input.graph The project graph in Nx's shape.
+ * @param {object[]} input.records The raw import-site records — the analysis
+ *   envelope's `imports` array verbatim (`../analysis/contract.md`), including
+ *   the `resolved: null` rows.
+ * @returns {object} The snapshot, ready for `serializeEvidenceSnapshot`.
+ * @throws {Error} naming the first piece of required structure that is missing
+ *   or malformed — a snapshot built over half-specified evidence would fail
+ *   later anyway, and farther from the cause.
+ */
+export function buildEvidenceSnapshot({
+  tool,
+  provenance,
+  provider,
+  policyFingerprint,
+  coverage,
+  graph,
+  records,
+}) {
+  if (!tool || typeof tool.name !== "string" || tool.name === "") {
+    throw new Error(
+      "archkeep: cannot build an evidence snapshot without a tool name — the snapshot must " +
+        "name the engine that captured it",
+    );
+  }
+  if (typeof tool.version !== "string" || tool.version === "") {
+    throw new Error(
+      "archkeep: cannot build an evidence snapshot without a tool version — a reader could not " +
+        "tell which engine revision produced the bytes",
+    );
+  }
+  if (typeof provider !== "string" || provider === "") {
+    throw new Error(
+      "archkeep: cannot build an evidence snapshot without a provider name — a delta run could " +
+        "not tell whether the baseline was read by the same project model it runs under",
+    );
+  }
+  if (typeof policyFingerprint !== "string" || policyFingerprint === "") {
+    throw new Error(
+      "archkeep: cannot build an evidence snapshot without a policy fingerprint — a delta run " +
+        "could not tell whether the boundary law changed between base and head",
+    );
+  }
+  const coverageProblems = describeCoverageProblems(coverage);
+  if (coverageProblems.length > 0) {
+    throw new Error(
+      "archkeep: cannot build an evidence snapshot — the coverage summary is malformed:\n  " +
+        coverageProblems.join("\n  "),
+    );
+  }
+  if (!graph || typeof graph !== "object" || typeof graph.nodes !== "object" || !graph.nodes) {
+    throw new Error(
+      "archkeep: cannot build an evidence snapshot without a project graph with a `nodes` map",
+    );
+  }
+  if (
+    typeof graph.dependencies !== "object" ||
+    graph.dependencies === null ||
+    Array.isArray(graph.dependencies)
+  ) {
+    throw new Error(
+      "archkeep: cannot build an evidence snapshot without the graph's `dependencies` map",
+    );
+  }
+  if (!Array.isArray(records)) {
+    throw new Error(
+      "archkeep: cannot build an evidence snapshot without the raw analysis records as an array — " +
+        "the records are what a delta run re-judges, and without them a baseline is a verdict " +
+        "that cannot be re-checked under changed law",
+    );
+  }
+  for (const [index, record] of records.entries()) {
+    if (record === null || typeof record !== "object" || Array.isArray(record)) {
+      throw new Error(
+        `archkeep: analysis record ${index} is ${describe(record)}, not an import-site object — ` +
+          "every record must carry the shape src/analysis/contract.md fixes",
+      );
+    }
+  }
+
+  const projects = buildProjects(graph.nodes).map((project) => {
+    // Re-attach the three rule-relevant fields `buildProjects` strips for the
+    // public graph contract. Each is attached only when the node DECLARES it —
+    // an absent field stays absent, because `evaluate()` treats absence as
+    // "this workspace declares none here", and inventing an empty value would
+    // be a second copy of that answer.
+    const data = graph.nodes[project.name]?.data ?? {};
+    /** @type {Record<string, unknown>} */
+    const extras = {};
+    if (data.mfeRemote !== undefined) extras.mfeRemote = data.mfeRemote;
+    if (Array.isArray(data.entryPoints)) {
+      extras.entryPoints = data.entryPoints.slice().sort(cmpString);
+    }
+    if (Array.isArray(data.declaredPackages)) {
+      extras.declaredPackages = data.declaredPackages.slice().sort(cmpString);
+    }
+    return { ...project, ...extras };
+  });
+
+  /** @type {Record<string, unknown>} */
+  const storedGraph = { projects, dependencies: buildDependencies(graph.dependencies) };
+  if (
+    graph.workspaceLayout !== undefined &&
+    graph.workspaceLayout !== null &&
+    typeof graph.workspaceLayout === "object"
+  ) {
+    storedGraph.workspaceLayout = graph.workspaceLayout;
+  }
+  if (Array.isArray(graph.exemptedFiles)) {
+    storedGraph.exemptedFiles = graph.exemptedFiles.slice().sort(cmpString);
+  }
+
+  return {
+    schemaVersion: EVIDENCE_SNAPSHOT_SCHEMA_VERSION,
+    tool: { name: tool.name, version: tool.version },
+    provider,
+    provenance,
+    policyFingerprint,
+    coverage: {
+      complete: coverage.complete,
+      analyzedFiles: coverage.analyzedFiles,
+      notAnalyzed: coverage.notAnalyzed,
+      blindSpots: coverage.blindSpots,
+    },
+    graph: storedGraph,
+    records,
+  };
+}
+
+/**
+ * Renders the snapshot as deterministic JSON text.
+ *
+ * Deterministic because `buildEvidenceSnapshot` constructs every key in a
+ * fixed order and sorts every array whose source does not guarantee order;
+ * two captures over one unchanged tree produce byte-identical files, which is
+ * what makes a plain `diff` of two baselines meaningful.
+ *
+ * @param {object} snapshot From `buildEvidenceSnapshot`.
+ * @returns {string} The JSON text, newline-terminated.
+ */
+export function serializeEvidenceSnapshot(snapshot) {
+  return `${JSON.stringify(snapshot, null, 2)}\n`;
+}
+
+/**
+ * Reads snapshot text from a path — the module's one filesystem seam, kept
+ * thin and injectable so tests and embedders drive validation without disk.
+ *
+ * @param {string} path Absolute path to the snapshot file.
+ * @param {{read?: (path: string) => string}} [io] Injectable read; defaults to
+ *   a UTF-8 `readFileSync`.
+ * @returns {object} Whatever `parseEvidenceSnapshot` returns for the text.
+ * @throws {Error} when the file cannot be read, naming the path and the cause,
+ *   and whatever `parseEvidenceSnapshot` throws.
+ */
+export function readEvidenceSnapshot(path, io = {}) {
+  const read = io.read ?? ((p) => readFileSync(p, "utf8"));
+  let text;
+  try {
+    text = read(path);
+  } catch (cause) {
+    throw new Error(
+      `archkeep: cannot read the evidence snapshot '${path}': ${cause?.message ?? cause}`,
+      { cause },
+    );
+  }
+  return parseEvidenceSnapshot(text, path);
+}
+
+/**
+ * Parses and validates snapshot text.
+ *
+ * Pure: text in, validated snapshot out. Every refusal names what is wrong —
+ * these become exit-3 ("could not complete") upstream, and a delta run that
+ * consumed a malformed baseline silently would classify against nothing while
+ * reporting a verdict.
+ *
+ * Refusals, each loud:
+ * - unreadable/malformed JSON — named with the path and the parse error;
+ * - a `schemaVersion` that is not the integer this format uses — a FUTURE
+ *   version refuses too: a reader that half-understood a newer format would
+ *   classify over evidence it misread;
+ * - any missing or malformed required section, all named together;
+ * - a baseline whose coverage is not complete. An incomplete base did not look
+ *   everywhere, so a head-only violation could exist unseen at base — classing
+ *   it "introduced" would fabricate a change the code may not contain. The
+ *   refusal names how many files went unanalyzed.
+ *
+ * What is deliberately NOT a refusal: dirty base provenance. A baseline from
+ * an uncommitted tree is weaker evidence, not unreadable evidence — the parsed
+ * snapshot exposes `provenance.dirty` so the renderer can say so loudly, and
+ * classification itself proceeds.
+ *
+ * @param {string} text The file contents.
+ * @param {string} path The path the text came from, for error messages.
+ * @returns {object} The validated snapshot.
+ * @throws {Error} on every condition above.
+ */
+export function parseEvidenceSnapshot(text, path) {
+  let parsed;
+  try {
+    parsed = JSON.parse(text);
+  } catch (cause) {
+    throw new Error(
+      `archkeep: the evidence snapshot '${path}' is not valid JSON: ${cause?.message ?? cause}`,
+      { cause },
+    );
+  }
+  if (parsed === null || typeof parsed !== "object" || Array.isArray(parsed)) {
+    throw new Error(
+      `archkeep: the evidence snapshot '${path}' must be a JSON object, got ${describe(parsed)}`,
+    );
+  }
+
+  const problems = [];
+  if (!Number.isInteger(parsed.schemaVersion)) {
+    problems.push(
+      `schemaVersion: must be the integer ${EVIDENCE_SNAPSHOT_SCHEMA_VERSION}, got ` +
+        `${describe(parsed.schemaVersion)}`,
+    );
+  } else if (parsed.schemaVersion !== EVIDENCE_SNAPSHOT_SCHEMA_VERSION) {
+    // Version 1 is the first version this format ever had, so a larger integer
+    // is a future format and a smaller one was never written by any release —
+    // the two refusals name different ways out.
+    const disposition =
+      parsed.schemaVersion > EVIDENCE_SNAPSHOT_SCHEMA_VERSION
+        ? "The file was written by a newer version of Archkeep; upgrade to read it."
+        : "No release of Archkeep ever wrote that version; re-capture the baseline.";
+    throw new Error(
+      `archkeep: the evidence snapshot '${path}' has schemaVersion ` +
+        `${parsed.schemaVersion}, which this tool does not understand — it understands ` +
+        `${EVIDENCE_SNAPSHOT_SCHEMA_VERSION}. ${disposition}`,
+    );
+  }
+
+  if (!isPlainObject(parsed.tool)) {
+    problems.push("tool: must be an object naming the capturing engine");
+  } else {
+    if (typeof parsed.tool.name !== "string" || parsed.tool.name === "") {
+      problems.push("tool.name: must be a non-empty string");
+    }
+    if (typeof parsed.tool.version !== "string" || parsed.tool.version === "") {
+      problems.push("tool.version: must be a non-empty string");
+    }
+  }
+
+  if (typeof parsed.provider !== "string" || parsed.provider === "") {
+    problems.push("provider: must be a non-empty string naming the project-model provider");
+  }
+
+  if (parsed.provenance !== null && !isPlainObject(parsed.provenance)) {
+    problems.push("provenance: must be an object ({commit, remote, dirty}) or null");
+  } else if (isPlainObject(parsed.provenance) && typeof parsed.provenance.commit !== "string") {
+    problems.push("provenance.commit: must be a string when provenance is present");
+  }
+
+  if (typeof parsed.policyFingerprint !== "string" || parsed.policyFingerprint === "") {
+    problems.push(
+      "policyFingerprint: must be a non-empty string — without it a delta run cannot tell " +
+        "whether the boundary law moved between base and head",
+    );
+  }
+
+  if (!isPlainObject(parsed.coverage)) {
+    problems.push(
+      "coverage: must be an object with the capture's coverage summary — a reader could not " +
+        "tell how complete the look behind the records was",
+    );
+  } else {
+    // Each shape problem already carries its own `coverage.`-prefixed name;
+    // the completeness reason is named first because it decides usability on
+    // its own.
+    const incomplete = incompleteBaselineCoverageReason(parsed.coverage);
+    if (incomplete) problems.push(incomplete);
+    problems.push(...describeCoverageProblems(parsed.coverage));
+  }
+
+  if (!isPlainObject(parsed.graph)) {
+    problems.push("graph: must be an object carrying projects and dependencies");
+  } else {
+    if (!Array.isArray(parsed.graph.projects)) {
+      problems.push("graph.projects: must be an array of project entries");
+    } else {
+      parsed.graph.projects.forEach((project, index) => {
+        if (!isPlainObject(project)) {
+          problems.push(`graph.projects[${index}]: must be an object`);
+          return;
+        }
+        if (typeof project.name !== "string" || project.name === "") {
+          problems.push(`graph.projects[${index}].name: must be a non-empty string`);
+        }
+        if (typeof project.root !== "string") {
+          problems.push(`graph.projects[${index}].root: must be a string`);
+        }
+      });
+    }
+    if (!Array.isArray(parsed.graph.dependencies)) {
+      problems.push("graph.dependencies: must be an array of edges");
+    } else {
+      parsed.graph.dependencies.forEach((edge, index) => {
+        if (!isPlainObject(edge)) {
+          problems.push(`graph.dependencies[${index}]: must be an object`);
+          return;
+        }
+        for (const field of ["source", "target", "type"]) {
+          if (typeof edge[field] !== "string") {
+            problems.push(`graph.dependencies[${index}].${field}: must be a string`);
+          }
+        }
+      });
+    }
+  }
+
+  if (!Array.isArray(parsed.records)) {
+    problems.push(
+      "records: must be an array of raw import-site records — without them the baseline holds " +
+        "nothing a delta run can re-judge under the current law",
+    );
+  } else {
+    parsed.records.forEach((record, index) => {
+      if (!isPlainObject(record)) {
+        problems.push(`records[${index}]: must be an object per ../analysis/contract.md`);
+        return;
+      }
+      if (typeof record.sourceFile !== "string" || record.sourceFile === "") {
+        problems.push(`records[${index}].sourceFile: must be a non-empty string`);
+      }
+      if (typeof record.specifier !== "string" || record.specifier === "") {
+        problems.push(`records[${index}].specifier: must be a non-empty string`);
+      }
+    });
+  }
+
+  if (problems.length > 0) {
+    throw new Error(
+      `archkeep: the evidence snapshot '${path}' is not a usable baseline:\n  ` +
+        problems.join("\n  "),
+    );
+  }
+  return parsed;
+}
+
+/**
+ * Whether the baseline's provider mismatches the current run's, as a reason
+ * string — `null` when they agree.
+ *
+ * Exported rather than folded into `parseEvidenceSnapshot` because the loader
+ * sees only the baseline: it cannot know what provider the current run uses.
+ * The check takes two records and decides, so the caller wires it where both
+ * sides are in hand.
+ *
+ * A mismatch means the two graphs were built by different project models, so
+ * structural differences may be provider artefacts rather than real changes —
+ * the same reasoning `./snapshot-meta.mjs` applies between graph snapshots.
+ *
+ * @param {string} baselineProvider What the snapshot recorded.
+ * @param {string} currentProvider What the current run resolved.
+ * @returns {string|null} The reason they conflict, or `null`.
+ */
+export function providerMismatch(baselineProvider, currentProvider) {
+  if (baselineProvider === currentProvider) return null;
+  return (
+    `the baseline was captured under the '${baselineProvider}' provider but this run is using ` +
+    `'${currentProvider}' — the two project models may attribute the same tree to different ` +
+    `projects, so structural differences may be provider artefacts rather than real changes`
+  );
+}
+
+/** Non-empty plain-object guard used across validation. */
+function isPlainObject(value) {
+  return value !== null && typeof value === "object" && !Array.isArray(value);
+}
+
+/** Describes a value for error messages without dumping it. */
+function describe(value) {
+  if (value === null) return "null";
+  if (Array.isArray(value)) return "an array";
+  return typeof value;
+}
+
+/** Plain lexicographic comparison — never localeCompare (byte-determinism). */
+function cmpString(a, b) {
+  return a < b ? -1 : a > b ? 1 : 0;
+}
+
+/**
+ * Coverage-shape problems, collected rather than thrown, so capture-time
+ * construction and load-time parsing share ONE statement of what coverage is —
+ * capture throws on the first, the parser folds every problem into its single
+ * refusal. Written once so the two cannot disagree about what coverage is.
+ *
+ * @param {object} coverage
+ * @returns {string[]} One entry per problem, empty when the shape is sound.
+ */
+function describeCoverageProblems(coverage) {
+  const problems = [];
+  if (!isPlainObject(coverage)) {
+    return [
+      "coverage: must be an object with the capture's coverage summary — a reader could not " +
+        "tell how complete the look behind the records was",
+    ];
+  }
+  if (typeof coverage.complete !== "boolean") {
+    problems.push(
+      "coverage.complete: must be a boolean — reading an unstated completeness as either " +
+        "answer would claim something the capture never said",
+    );
+  }
+  if (typeof coverage.analyzedFiles !== "number" || !Number.isFinite(coverage.analyzedFiles)) {
+    problems.push("coverage.analyzedFiles: must be a finite number");
+  }
+  if (!Array.isArray(coverage.notAnalyzed) || !Array.isArray(coverage.blindSpots)) {
+    problems.push(
+      "coverage.notAnalyzed and coverage.blindSpots: both must be arrays — both are always " +
+        "arrays in the analysis envelope, and a consumer iterates them without checking",
+    );
+  }
+  return problems;
+}
+
+/**
+ * The reason a coverage summary cannot serve as a delta BASELINE, or `null`.
+ *
+ * Load-side only, deliberately: capture may honestly record an incomplete look
+ * (the evidence includes WHICH files went unanalyzed), but consuming one as a
+ * comparison base would fabricate classifications — a violation living in a
+ * file the base never looked at reads as newly introduced at head even if it
+ * predates the run. The refusal names how many files went unanalyzed.
+ *
+ * @param {object} coverage
+ * @returns {string|null}
+ */
+function incompleteBaselineCoverageReason(coverage) {
+  if (
+    coverage &&
+    typeof coverage === "object" &&
+    !Array.isArray(coverage) &&
+    coverage.complete === false &&
+    Array.isArray(coverage.notAnalyzed)
+  ) {
+    return (
+      `coverage.complete: the baseline's coverage is not complete — ` +
+      `${coverage.notAnalyzed.length} file(s) could not be analyzed at capture time, so a ` +
+      `violation living there would be misread as newly introduced at head`
+    );
+  }
+  return null;
+}