@ecoma-io/archkeep 0.16.1 → 0.18.0

package/src/governance/evolution-event.mjs

@@ -0,0 +1,470 @@
+/**
+ * The canonical evolution event — one record every evolution command emits,
+ * with one stable identity, one classification vocabulary and one idempotency
+ * rule. `docs/concepts/evolution.md` is the human-facing model; this module is
+ * its implementation, and the two must not disagree.
+ *
+ * The whole module is pure: no clock, no filesystem, no injected state. That
+ * is the load-bearing half of the design. An event's `id`/`dedupeKey` are
+ * derived from `{base, head, declarationDigest}` and nothing else — never from
+ * `recordedAt`, `notes`, or `provenance` — so re-running the same transition
+ * produces the same event and the store can prove idempotency instead of
+ * guessing it. `declarationDigest` covers only the DECLARATIVE parts of a
+ * change-intent (`{version, base, projects, edges, constraints}`); free-prose
+ * `summary` is excluded because prose is not semantics: two runs whose
+ * summary was re-worded must produce the same digest, or the digest would be
+ * a function of narration rather than of the declared change.
+ *
+ * `classifyEvolution(input)` is the one home of the classification predicates
+ * (design §2): CHANGE, DRIFT, VIOLATION, REPAIR, DECISION_CHANGE. Each class
+ * is a fact about the input, multiple classes are allowed, and the output is
+ * sorted. The invariant this module exists to hold is the repository's own:
+ * **an empty result is a claim, not a shrug.** An unclassifiable item — an
+ * unknown delta entry, one-sided metadata, a verdict that could not be
+ * determined — is never folded into a clean result: it raises a `notes[]`
+ * disclosure, and where the event carries a verdict-relevant unknown the
+ * disposition is `no-verdict`, never a fabricated `accepted`/`rejected`.
+ * `[]` classifications appear only for a fully comparable, unchanged pair,
+ * and the event says so in notes.
+ */
+
+import { createHash } from "node:crypto";
+
+import { canonicalizeJson } from "../canonical.mjs";
+
+/** The version every event record carries. A different value is a store error. */
+export const EVOLUTION_EVENT_SCHEMA_VERSION = 1;
+
+/** The classification vocabulary, one fact per class, sorted lexicographically. */
+export const EVENT_CLASSIFICATIONS = Object.freeze([
+  "CHANGE",
+  "DRIFT",
+  "VIOLATION",
+  "REPAIR",
+  "DECISION_CHANGE",
+]);
+
+/** The disposition vocabulary every event may carry. */
+export const EVENT_DISPOSITIONS = Object.freeze(["accepted", "rejected", "no-verdict"]);
+
+/**
+ * The canonical tuple an event's identity is built on — the ONE definition,
+ * shared by `eventId` (which hashes it) and `eventDedupeKey` (which is it), so
+ * the store cannot dedupe on one spelling of the tuple while the id is another.
+ *
+ * `declarationDigest` participates only when the event carries a declaration:
+ * a transition-kind event (history/evolution) has none, and its absence is
+ * part of the tuple — an event with a declaration is a different event from
+ * the same base/head without one, and must not collide with it.
+ *
+ * @param {{base: object, head: object, declaration?: {digest?: string}}} event
+ * @returns {string} The canonical serialization of the tuple.
+ */
+export function eventDedupeKey(event) {
+  return canonicalizeJson({
+    base: event.base,
+    head: event.head,
+    declarationDigest: event.declaration?.digest,
+  });
+}
+
+/**
+ * The stable event identity: `sha256` of the canonical tuple. Never depends on
+ * `recordedAt`, `notes`, `provenance`, or any wall clock — see the module
+ * header for why that exclusion is the idempotency guarantee.
+ *
+ * @param {{base: object, head: object, declaration?: {digest?: string}}} event
+ * @returns {string} 64-hex-char SHA-256.
+ */
+export function eventId(event) {
+  return createHash("sha256").update(eventDedupeKey(event)).digest("hex");
+}
+
+/**
+ * The digest of a normalized change-intent's DECLARATIVE parts only:
+ * `{version, base, projects, edges, constraints}`. The prose `summary` is
+ * excluded by construction — this object never references it, so a re-worded
+ * summary cannot change the digest (idempotency: prose is not semantics).
+ *
+ * The input is the same normalized shape `commands/change-intent.mjs`'s
+ * `parseChangeIntent` returns; the digest is built from the whole declarative
+ * sections, so any declarative difference — a project row, an edge row, a
+ * constraint key, the base commit — changes the digest.
+ *
+ * @param {{version: string, base: object, projects: object, edges: object,
+ *   constraints: object, summary?: string}} intent A normalized change-intent.
+ *   `summary` is accepted (it rides on the normalized intent) and deliberately
+ *   ignored — prose is not semantics.
+ * @returns {string} The canonical serialization of the declarative parts.
+ */
+export function declarationDigest(intent) {
+  return canonicalizeJson({
+    version: intent.version,
+    base: intent.base,
+    projects: intent.projects,
+    edges: intent.edges,
+    constraints: intent.constraints,
+  });
+}
+
+/**
+ * @typedef {object} EvolutionEvidence
+ * @property {{projects?: {added: string[], removed: string[], changed: string[]},
+ *   edges?: {added: string[], removed: string[]},
+ *   policyChanged?: boolean|null, policyOneSided?: boolean,
+ *   provenanceChanged?: boolean|null}} [observed]
+ *   The structural diff between base and head: project names and edge identity
+ *   strings (source,target,type) that were added, removed, or changed. Empty
+ *   by default. `policyChanged` — whether the policy fingerprint changed
+ *   between base and head; `null` is "could not be compared": exactly one side
+ *   records the policy (`policyOneSided: true`) or neither does
+ *   (both-absent). `true` is a disclosure, never a refusal. `policyOneSided`
+ *   is the one-sided mirror from `commands/history.mjs` — an input fact, never
+ *   derived from `policyChanged === null`, because both-sides-absent also
+ *   yields `null` while remaining comparable. `provenanceChanged` — whether
+ *   the commit record advanced between base and head; it is `true` only when
+ *   both sides record provenance and the commits differ.
+ * @property {boolean} [codeDrift]
+ *   Whether provenance advanced with no architectural or policy change. The
+ *   caller computes it under `commands/history.mjs`'s discipline — only when
+ *   the architecture did not move, the policy was ACTUALLY compared, and the
+ *   policy did not change; a caller that passes `true` alongside a
+ *   `policyChanged` of `null` gets a loud note and no DRIFT, never a guess.
+ * @property {{introduced?: {id: string, waived: boolean}[],
+ *   resolved?: string[], unknown?: {id: string, reason: string}[]}} [violations]
+ *   The delta classification of violations: introduced (with each entry's
+ *   waiver state), resolved, and unknown. Unknown entries are the
+ *   verdict-relevant unclassifiable items — they force `no-verdict`.
+ * @property {{id: string, verdict: "pass"|"fail"|"unknown"}[]} [declaredConstraints]
+ *   Declared constraint verdicts. `fail` is a VIOLATION; any other value than
+ *   `pass`/`fail` is an undeterminable verdict — note + `no-verdict`.
+ * @property {{id: string,
+ *   verdict: "matched"|"undeclared"|"unfulfilled"|"unproven"}[]} [declaredIntentRows]
+ *   Declared change-intent rows and the head's verdict on each (the
+ *   `matched|undeclared|unfulfilled|unproven` vocabulary). `undeclared` and
+ *   `unfulfilled` are DRIFT and reject; `unproven` is note + `no-verdict`.
+ * @property {string[]} [driftFindingsResolved]
+ *   Drift findings that no longer exist at head — a REPAIR signal.
+ * @property {string[]} [debtResolved]
+ *   Active debt entries closed between base and head — a REPAIR signal.
+ * @property {{records: {id: string, status: string, supersedes: string[]}[]}|null} [adrBase]
+ *   The ADR registry at base, as `{records: …}` (the shape
+ *   `adr-registry.mjs`'s `loadAdrRegistry` returns), or `null` when the side
+ *   does not record a decision registry. `undefined` = the input carried no
+ *   decision evidence at all (not one-sided).
+ * @property {{records: {id: string, status: string, supersedes: string[]}[]}|null} [adrHead]
+ *   The ADR registry at head, or `null` when absent. Exactly one side `null`
+ *   is the one-sided case: DECISION_CHANGE is NOT asserted and a note is
+ *   added.
+ */
+
+/**
+ * The classification result.
+ *
+ * @typedef {object} EvolutionClassification
+ * @property {string[]} classifications Sorted; a subset of
+ *   `EVENT_CLASSIFICATIONS`. `[]` only for a fully comparable, unchanged pair
+ *   (with a note saying so) — never for an input that could not be classified.
+ * @property {"accepted"|"rejected"|"no-verdict"} disposition
+ *   `no-verdict` on any verdict-relevant unknown; `rejected` on introduced
+ *   non-waived violations, a declared constraint `fail`, or undeclared /
+ *   unfulfilled declared-intent rows; `accepted` otherwise.
+ * @property {string[]} notes Disclosure notes — one-sided metadata, unknown
+ *   entries, the "fully comparable, unchanged" statement, the policy-change
+ *   disclosure.
+ * @property {{projects: string[], boundaries: string[], constraints: string[],
+ *   decisions: string[]}} affected Identity strings only, each sorted and
+ *   de-duplicated.
+ */
+
+/**
+ * Classifies one evolution event from its evidence signals — deterministic
+ * and pure (no clock, no fs). Each predicate is a fact about the input,
+ * never an inference (design §2):
+ *
+ * | Class | Predicate (all must hold) |
+ * |---|---|
+ * | CHANGE | structural diff between base/head non-empty |
+ * | DRIFT | `codeDrift` signal, OR declared intent rows `undeclared`/`unfulfilled`
+ *   by the head; and NOT merely a policy-only transition; and no
+ *   DECISION_CHANGE asserted (a supersession is DECISION_CHANGE, never DRIFT) |
+ * | VIOLATION | introduced violations non-empty with ≥1 not-waived, OR a declared
+ *   constraint verdict `fail` |
+ * | REPAIR | resolved violations non-empty, OR drift findings resolved, OR active
+ *   debt entries closed |
+ * | DECISION_CHANGE | same ADR id with a different status between base/head, or a
+ *   new `supersedes` relation. REQUIRES both sides' registries — either side
+ *   absent (`null`) ⇒ NOT asserted, note added |
+ *
+ * The `affected` identities are derived from the same signals, never from a
+ * second opinion: changed project names, changed edge identity strings, the
+ * constraint/intent rows whose verdict was not `pass`/`matched`, and the ADR
+ * ids whose lineage moved.
+ *
+ * @param {EvolutionEvidence} [input]
+ * @returns {EvolutionClassification}
+ */
+export function classifyEvolution(input = {}) {
+  const observed = input.observed ?? {};
+  const projects = observed.projects ?? { added: [], removed: [], changed: [] };
+  const edges = observed.edges ?? { added: [], removed: [] };
+  const addedProjects = projects.added ?? [];
+  const removedProjects = projects.removed ?? [];
+  const changedProjects = projects.changed ?? [];
+  const addedEdges = edges.added ?? [];
+  const removedEdges = edges.removed ?? [];
+  const structureChanged =
+    addedProjects.length +
+      removedProjects.length +
+      changedProjects.length +
+      addedEdges.length +
+      removedEdges.length >
+    0;
+
+  const codeDrift = input.codeDrift === true;
+  const policyChanged = observed.policyChanged ?? false;
+  // One-sided is an input fact (`snapshot-meta`'s `policyOneSided`), never
+  // derived from `policyChanged === null`: both-absent also yields `null` but
+  // is comparable, and deriving one-sided from it would disclose a refusal
+  // that is not real.
+  const policyOneSided = observed.policyOneSided === true;
+  // Provenance actually advanced (both sides record a commit and they
+  // differ). Pair an unverifiable policy with advancing provenance and the
+  // transition carried real code motion that cannot be classified — never a
+  // "fully comparable, unchanged pair".
+  const provenanceAdvanced = observed.provenanceChanged === true;
+
+  const violations = input.violations ?? {};
+  const introduced = violations.introduced ?? [];
+  const resolved = violations.resolved ?? [];
+  const unknown = violations.unknown ?? [];
+  const declaredConstraints = input.declaredConstraints ?? [];
+  const declaredIntentRows = input.declaredIntentRows ?? [];
+  const driftFindingsResolved = input.driftFindingsResolved ?? [];
+  const debtResolved = input.debtResolved ?? [];
+
+  const adrBase = input.adrBase;
+  const adrHead = input.adrHead;
+  // One-sided mirror of history's policyOneSided: exactly ONE side carries a
+  // registry. Both sides absent ("no decision evidence supplied", like
+  // snapshots that record no fingerprint on either side) is comparable —
+  // nothing was asserted, and there is nothing to disclose. Either side
+  // `null` (the side does not record one) while the other carries the
+  // registry is "could not be compared".
+  const decisionOneSided = (adrBase == null) !== (adrHead == null);
+
+  const notes = [];
+  const classifications = new Set();
+  const constraintIds = [];
+  const decisionIds = [];
+
+  if (structureChanged) {
+    classifications.add("CHANGE");
+  }
+
+  // Disclosures first, so every note is in the record even when a predicate
+  // below has nothing to add — a transition's notes must not depend on its
+  // classes. The wording mirrors `commands/history.mjs`'s `classifyTransition`
+  // for the same signals, so the two surfaces read the same facts the same way.
+  if (policyChanged === true) {
+    notes.push(
+      "policy (the declared architectural intent) changed between these snapshots — " +
+        "the boundary law differs even though the graph may not",
+    );
+  }
+  if (policyOneSided) {
+    notes.push(
+      "policy (the declared architectural intent) could not be compared — one side of " +
+        "the transition records the boundary law and the other does not",
+    );
+  }
+  if (observed.policyChanged === null && provenanceAdvanced) {
+    notes.push(
+      "policy (the declared architectural intent) could not be compared — neither side " +
+        "records the boundary law while the provenance advanced, so the pair cannot be " +
+        "classified (code motion may be hidden)",
+    );
+  }
+  if (decisionOneSided) {
+    notes.push(
+      "the decision registry (docs/adr) could not be compared — one side of the " +
+        "transition records it and the other does not",
+    );
+  }
+
+  // Verdict-relevant unknowns: an entry the input cannot classify is NEVER
+  // folded into a clean result. Each one is disclosed, and any of them makes
+  // the disposition `no-verdict` (never a fabricated accepted/rejected).
+  let verdictRelevantUnknown = false;
+  for (const entry of unknown) {
+    verdictRelevantUnknown = true;
+    notes.push(
+      `delta entry '${entry.id}' could not be classified` +
+        (typeof entry.reason === "string" && entry.reason !== "" ? ` — ${entry.reason}` : ""),
+    );
+  }
+  for (const row of declaredConstraints) {
+    constraintIds.push(row.id);
+    if (row.verdict !== "pass" && row.verdict !== "fail") {
+      verdictRelevantUnknown = true;
+      notes.push(
+        `declared constraint '${row.id}' verdict '${row.verdict}' could not be determined`,
+      );
+    }
+  }
+  for (const row of declaredIntentRows) {
+    if (
+      row.verdict !== "matched" &&
+      row.verdict !== "undeclared" &&
+      row.verdict !== "unfulfilled" &&
+      row.verdict !== "unproven"
+    ) {
+      verdictRelevantUnknown = true;
+      notes.push(`declared intent row '${row.id}' carries an unknown verdict '${row.verdict}'`);
+    } else if (row.verdict === "unproven") {
+      verdictRelevantUnknown = true;
+      notes.push(
+        `declared intent row '${row.id}' is unproven — its verdict could not be determined`,
+      );
+    }
+    if (row.verdict !== "matched") {
+      constraintIds.push(row.id);
+    }
+  }
+
+  // VIOLATION — introduced non-waived entries are the violations that count; a
+  // waived introduction is a disclosed, non-gating fact, never a violation —
+  // and an introduction whose every entry is waived is disclosed too, so the
+  // empty classification below never reads as "nothing moved".
+  const introducedNotWaived = introduced.filter((entry) => entry.waived !== true);
+  const failedConstraints = declaredConstraints.filter((row) => row.verdict === "fail");
+  if (introducedNotWaived.length > 0 || failedConstraints.length > 0) {
+    classifications.add("VIOLATION");
+  }
+  if (introduced.length > 0 && introducedNotWaived.length === 0) {
+    notes.push("introduced violations are all waived — not classified as VIOLATION");
+  }
+
+  // REPAIR — resolved violations, resolved drift findings, or closed debt.
+  if (resolved.length > 0 || driftFindingsResolved.length > 0 || debtResolved.length > 0) {
+    classifications.add("REPAIR");
+  }
+
+  // DECISION_CHANGE — the ADR lineage moved between base and head. Both sides'
+  // registries are REQUIRED: either side absent is the one-sided case, and
+  // "could not be compared" is disclosed, never asserted as a change.
+  if (!decisionOneSided) {
+    const baseRecords = adrBase?.records ?? [];
+    const headRecords = adrHead?.records ?? [];
+    const baseStatus = new Map(baseRecords.map((record) => [record.id, record.status]));
+    const baseSupersedes = new Map(
+      baseRecords.map((record) => [record.id, new Set(record.supersedes ?? [])]),
+    );
+    for (const record of headRecords) {
+      if (baseStatus.has(record.id) && baseStatus.get(record.id) !== record.status) {
+        decisionIds.push(record.id);
+      }
+      const relationsAtBase = baseSupersedes.get(record.id) ?? new Set();
+      for (const target of record.supersedes ?? []) {
+        if (!relationsAtBase.has(target)) {
+          decisionIds.push(record.id, target);
+        }
+      }
+    }
+    if (decisionIds.length > 0) {
+      classifications.add("DECISION_CHANGE");
+    }
+  }
+
+  // DRIFT — the one predicate with a guard and an exclusion. The guard: not
+  // merely a policy-only transition (a policy change with no other signal is
+  // disclosed, never DRIFT). The exclusion: a supersession between two
+  // registry states is DECISION_CHANGE, never DRIFT (design §9), so the code
+  // drift signal is not ALSO classified DRIFT when the lineage moved.
+  if (!classifications.has("DECISION_CHANGE")) {
+    const otherSignals =
+      structureChanged ||
+      codeDrift ||
+      introduced.length > 0 ||
+      resolved.length > 0 ||
+      unknown.length > 0 ||
+      driftFindingsResolved.length > 0 ||
+      debtResolved.length > 0 ||
+      declaredConstraints.some((row) => row.verdict !== "pass") ||
+      declaredIntentRows.some((row) => row.verdict !== "matched");
+    const merelyPolicyOnly = policyChanged === true && !otherSignals;
+    const intentViolated = declaredIntentRows.filter(
+      (row) => row.verdict === "undeclared" || row.verdict === "unfulfilled",
+    );
+    let drifted = intentViolated.length > 0;
+    if (codeDrift) {
+      if (policyOneSided) {
+        notes.push("code drift cannot be asserted — the policy change could not be compared");
+      } else if (!merelyPolicyOnly) {
+        drifted = true;
+      }
+    }
+    if (drifted) {
+      classifications.add("DRIFT");
+      for (const row of intentViolated) {
+        constraintIds.push(row.id);
+      }
+    }
+  }
+
+  // The disposition: `no-verdict` on any verdict-relevant unknown (never a
+  // fabricated accepted/rejected), `rejected` on introduced non-waived
+  // violations, a declared constraint `fail`, or undeclared / unfulfilled
+  // declared-intent rows, `accepted` otherwise — including the fully
+  // comparable, unchanged pair.
+  /** @type {"accepted"|"rejected"|"no-verdict"} */
+  let disposition = "accepted";
+  if (verdictRelevantUnknown) {
+    disposition = "no-verdict";
+  } else if (
+    introducedNotWaived.length > 0 ||
+    failedConstraints.length > 0 ||
+    declaredIntentRows.some((row) => row.verdict === "undeclared" || row.verdict === "unfulfilled")
+  ) {
+    disposition = "rejected";
+  }
+
+  const classificationsList = [...classifications].sort();
+
+  // **An empty result is a claim.** `[]` classifications appear ONLY for a
+  // fully comparable, unchanged pair, and the event says so; an input that
+  // could not be fully classified never reads as a clean "nothing happened".
+  // Each other empty-list case carries its own disclosure (the one-sided
+  // notes, the waiver note, the unknown-entry notes), so no branch of this
+  // block produces a silent `[]`.
+  if (classificationsList.length === 0) {
+    const signaledButUnclassified =
+      introduced.length > 0 ||
+      resolved.length > 0 ||
+      driftFindingsResolved.length > 0 ||
+      debtResolved.length > 0;
+    if (
+      !verdictRelevantUnknown &&
+      !signaledButUnclassified &&
+      !policyOneSided &&
+      !decisionOneSided &&
+      !(observed.policyChanged === null && provenanceAdvanced)
+    ) {
+      notes.push("a fully comparable, unchanged pair — no classification applies");
+    }
+  }
+
+  const uniqueSorted = (values) => [...new Set(values)].sort();
+
+  return {
+    classifications: classificationsList,
+    disposition,
+    notes,
+    affected: {
+      projects: uniqueSorted([...addedProjects, ...removedProjects, ...changedProjects]),
+      boundaries: uniqueSorted([...addedEdges, ...removedEdges]),
+      constraints: uniqueSorted(constraintIds),
+      decisions: uniqueSorted(decisionIds),
+    },
+  };
+}