@sdd-method/maintain-cli 0.0.0-stage → 0.1.0

package/dist/lib/catalogue/baseline-check.js

@@ -0,0 +1,151 @@
+/**
+ * `catalogue baseline-check` — compares emitted findings against a repo's
+ * findings baseline BY IDENTITY (M-ADR 0132 §9), and reports the four outcomes
+ * a re-take needs to be reviewable: unbaselined, reworded, closed, unchanged.
+ *
+ * WHY THIS IS A COMMAND AND NOT A DISCIPLINE. The comparison was previously
+ * described in prose and performed per repo, which is why the rewording class
+ * (§9.1) went undetected twice on the same finding family: every repo
+ * implemented "is this line in the file" independently and correctly, and the
+ * defect was in the question, not the implementations.
+ *
+ * It reads findings rather than producing them, so it composes with whatever
+ * produced them and cannot disagree with a separate run of the builder:
+ *
+ *   sdd-cli catalogue validate --coverage > findings.txt
+ *   sdd-cli catalogue baseline-check --findings findings.txt
+ *
+ * ONLY GATE-ACTING CATEGORIES ARE COMPARED (`orphan`, `divergence` — M-ADR 0132
+ * §4.3). Filtering happens here, visibly, and the count of what was filtered is
+ * printed: a comparator that silently dropped categories would make an empty
+ * result ambiguous between "nothing gate-acting" and "nothing at all".
+ */
+import { readFile } from "node:fs/promises";
+import { join } from "node:path";
+import { compareToBaseline } from "./finding-identity.js";
+/** The categories a baseline governs, per M-ADR 0132 §4.3. */
+const GATE_ACTING = new Set(["orphan", "divergence"]);
+export const DEFAULT_BASELINE_REL = join("orchestration", "catalogue-findings-baseline.txt");
+const FINDING_LINE = /^\[finding\] (\w+)(?: \[error\])?[ (:]/;
+/**
+ * Findings from a stream of mixed output. Non-finding lines (metrics, headers,
+ * blank lines, a baseline file's `#` provenance) are not findings and are
+ * dropped without comment; a `[finding]` line of a non-gate-acting category is
+ * dropped and COUNTED, because that number is the difference between "nothing
+ * to compare" and "nothing was comparable".
+ */
+export function extractGateActing(text) {
+    const lines = [];
+    let nonGateActing = 0;
+    for (const raw of text.split("\n")) {
+        const line = raw.trimEnd();
+        const m = FINDING_LINE.exec(line);
+        if (!m)
+            continue;
+        if (GATE_ACTING.has(m[1]))
+            lines.push(line);
+        else
+            nonGateActing++;
+    }
+    return { lines, nonGateActing };
+}
+export async function runBaselineCheck(options, write) {
+    const root = options.sddPath ?? process.cwd();
+    const baselinePath = options.baseline ?? join(root, DEFAULT_BASELINE_REL);
+    let findingsText;
+    if (options.findings !== undefined) {
+        findingsText = await readFile(options.findings, "utf-8");
+    }
+    else if (options.stdin) {
+        findingsText = await options.stdin();
+    }
+    else {
+        write("[error] baseline-check: no findings source (--findings <path>, or pipe them in)\n");
+        return { exitCode: 2 };
+    }
+    // An input with no lines at all is not "zero findings": every real findings
+    // stream (build, coverage, validate output) carries at least its own metric
+    // and verdict lines. The default source is stdin, so an empty pipe arrives
+    // here as nothing — refused, never compared (empty-set sweep, R2).
+    if (findingsText.trim() === "") {
+        write(`[error] baseline-check: the findings stream is EMPTY — no lines of any kind. ` +
+            `This is not zero findings; it is no measurement (an empty pipe, or the wrong ` +
+            `file). Pass --findings <path> to the captured stream\n`);
+        return { exitCode: 2 };
+    }
+    let baselineText;
+    try {
+        baselineText = await readFile(baselinePath, "utf-8");
+    }
+    catch (err) {
+        if (err.code !== "ENOENT")
+            throw err;
+        // A MISSING baseline is not an empty one. Under §9.3 the canonical zero is
+        // an empty-but-present file, precisely so that "measured, and there were
+        // none" is distinguishable from "never measured" — so this reports rather
+        // than treating every live finding as unbaselined.
+        write(`[error] baseline-check: no baseline at ${baselinePath} — a missing file is not a ` +
+            `recorded zero (M-ADR 0132 §9.3); create it empty, with its provenance header, ` +
+            `to record that findings were measured and there were none\n`);
+        return { exitCode: 2 };
+    }
+    const live = extractGateActing(findingsText);
+    const base = extractGateActing(baselineText);
+    const cmp = compareToBaseline(live.lines, base.lines);
+    write(`[metric] baseline_check: ${live.lines.length} gate-acting finding(s) against ` +
+        `${base.lines.length} baseline entr(ies) — ${cmp.unchanged} unchanged, ` +
+        `${cmp.reworded.length} reworded, ${cmp.unbaselined.length} unbaselined, ` +
+        `${cmp.stale.length} closed; ${live.nonGateActing} non-gate-acting finding(s) not compared\n`);
+    // ⚠ THE WHOLE-BASELINE-CLOSED GUARD. An all-closed result against a non-empty
+    // baseline is far more likely to mean the comparison read the WRONG STREAM than
+    // that an estate fixed everything at once, and until now it rendered as a clean
+    // green number: "0 gate-acting findings, 215 closed" reads as success and is
+    // indistinguishable from a gate comparing against nothing.
+    //
+    // Raised by an adopter estate that hit it on a 215-entry baseline: its
+    // baseline had been captured from one stream/tier and was being compared
+    // against another, so every entry reported closed and the regression gate
+    // passed unconditionally. The `--findings` default is stdin, so an empty pipe
+    // reaches here as legitimately-zero findings — the same estate had already
+    // been caught once by a bare invocation reporting 832 closed.
+    //
+    // The condition is deliberately narrow: live gate-acting findings ZERO, baseline
+    // non-empty, and EVERY entry closed. That triple cannot arise from ordinary
+    // remediation without also being worth a human look.
+    const allClosed = live.lines.length === 0 && base.lines.length > 0 && cmp.stale.length === base.lines.length;
+    if (allClosed) {
+        write(`[finding] divergence: baseline_check compared ZERO live gate-acting findings ` +
+            `against ${base.lines.length} baseline entr(ies) and reported ALL of them closed. ` +
+            `This is reported as a DEFECT rather than a pass: an all-closed baseline is far ` +
+            `more likely to mean the findings stream was empty or was the wrong one than that ` +
+            `every finding was fixed at once. Check that --findings received the stream the ` +
+            `baseline was CAPTURED from (note the default is stdin, so an empty pipe arrives ` +
+            `here as zero findings), and that --tier matches. Re-taking the baseline against ` +
+            `this comparison would bake the empty stream in as the new truth and destroy the ` +
+            `${base.lines.length}-entry record.\n`);
+    }
+    // Reported even at zero. While it is above zero those entries are matched on
+    // whole text and remain exposed to the rewording class; silence here would
+    // imply a protection the repo has not got.
+    write(`[metric] baseline_check.legacy_keyed: ${cmp.legacyKeyed} baseline entr(ies) carry no ` +
+        `identity and are matched on their whole line — re-take them as their emission ` +
+        `sites migrate (M-ADR 0132 §9.2)\n`);
+    for (const r of cmp.reworded) {
+        // Named, not counted only: a re-take must be able to show which line
+        // replaced which, and "N reworded" is not reviewable evidence.
+        write(`[reworded] ${r.baseline}\n[reworded]   -> ${r.live}\n`);
+    }
+    for (const s of cmp.stale)
+        write(`[closed] ${s}\n`);
+    for (const u of cmp.unbaselined)
+        write(`[unbaselined] ${u}\n`);
+    // Only genuinely unbaselined findings fail. A reworded finding is the case
+    // this whole mechanism exists to stop failing, and a closed one is good news
+    // that a re-take tidies — neither is a regression.
+    // The all-closed guard FAILS the gate (empty-set sweep, R2, ruled 2026-09-29):
+    // it used to print "a DEFECT rather than a pass" and exit 0, so the regression
+    // gate still passed on exactly the empty/wrong stream it was written to catch.
+    // A genuinely all-fixed estate re-takes its baseline once, deliberately.
+    return { exitCode: cmp.unbaselined.length > 0 || allClosed ? 1 : 0 };
+}
+//# sourceMappingURL=baseline-check.js.map