@ybotok/loam 0.1.0-beta.3 → 0.2.0-alpha.13

package/dist/core/provenance/stamp.js

@@ -0,0 +1,220 @@
+/**
+ * The stamp vocabulary: what `loam vouch` computes for a frontmatter stamp and
+ * what a later check reads back out of one — `sources_digest` and the per-file
+ * `sources_files` index beside it, and the local calendar day `last_verified`
+ * carries. The document's own `content_digest` lived here too; it moved to
+ * core/document/frontmatter.ts, beside the body cut it hashes, when the fleet
+ * enumeration needed it (`contentDigest`'s docblock says why).
+ */
+import { createHash } from "node:crypto";
+import { readFile } from "node:fs/promises";
+import { collectSources } from "./walk.js";
+/**
+ * The local calendar day, for the two commands that date a stamp. A vouch or a
+ * verification record is a person saying "today I read this", so it is their
+ * date, not UTC's — `toISOString` files an evening in the Americas under
+ * tomorrow. It sits here because this module already owns the stamp vocabulary
+ * the date is written into.
+ */
+export function today(now) {
+    const pad = (n) => String(n).padStart(2, "0");
+    return `${now.getFullYear()}-${pad(now.getMonth() + 1)}-${pad(now.getDate())}`;
+}
+/* ------------------------------------------------------------------ */
+/* The sources digest                                                  */
+/* ------------------------------------------------------------------ */
+/** How much of the sha256 is written into the document. */
+const DIGEST_LENGTH = 16;
+/** How many source files the digest reads at once — see `sourcesDigest`. */
+const DIGEST_READ_BATCH = 32;
+/**
+ * Digest the CONTENT of the files `sources` names. The recipe is spelled out
+ * because the value is written into documents, and anything reading them has to
+ * be able to reproduce it:
+ *
+ *   1. expand each entry to repo-relative file paths (`/` separators) — a file
+ *      is itself, a directory everything beneath it, minus what the repository
+ *      does not consider its own (see collectSources) — sorted and de-duplicated;
+ *   2. per file, `sha256(bytes)`;
+ *   3. feed `<path>\0<hex>\n` for each file, in that order, into an outer
+ *      sha256;
+ *   4. keep the first 16 hex characters.
+ *
+ * Content, not mtime: git does not preserve modification times, so after a
+ * fresh clone every file looks changed and the check would be false positives
+ * end to end. Bytes survive the clone. Hashing the path alongside the content
+ * means a rename registers, which is what a reader of the doc would want.
+ *
+ * 64 bits is a change detector, not a seal — it answers "did this move?", and
+ * an adversary who wants a collision can have one.
+ */
+export async function sourcesDigest(repoDir, sources) {
+    const { found, skipped } = await collectSources(repoDir, sources);
+    const outer = createHash("sha256");
+    const index = [];
+    // Read in fixed batches, hash in array order. `found` is already sorted and
+    // every file's own sha depends on nothing but its own bytes, so the sequence
+    // fed to `outer` — and therefore the digest — is unchanged by construction;
+    // only the waiting overlaps. A `sources: [src/]` over a real service repo is
+    // thousands of files, and one-await-per-file spent nearly all of its time
+    // idle. The batch is bounded rather than a single `Promise.all` over the
+    // whole list because an unbounded fan-out over a source tree runs the process
+    // out of file descriptors, which would turn a stamp into an EMFILE.
+    for (let start = 0; start < found.length; start += DIGEST_READ_BATCH) {
+        const hashed = await Promise.all(found.slice(start, start + DIGEST_READ_BATCH).map(async (file) => ({
+            rel: file.rel,
+            content: createHash("sha256").update(await readFile(file.abs)).digest("hex"),
+        })));
+        for (const { rel, content } of hashed) {
+            outer.update(`${rel}\0${content}\n`);
+            index.push({ path: rel, sha: content.slice(0, DIGEST_LENGTH) });
+        }
+    }
+    // `empty` is deliberately not set here: it is a sentence about a named
+    // document, and a digest has no name to put in it. Callers that need it say
+    // emptySourcesMessage(label, sources) — the same words, from one definition.
+    return { digest: outer.digest("hex").slice(0, DIGEST_LENGTH), files: index.map((e) => e.path), index, skipped };
+}
+/**
+ * What a `sources` list actually covers — the digest's file set without the
+ * hashing.
+ *
+ * Exported because two commands must give the SAME answer to "does this list
+ * cover anything?". `loam vouch` refuses to stamp an expansion that covers no
+ * files (a digest over nothing never changes, so the stamp would read as
+ * current forever), and until `loam validate` could say the same thing in the
+ * same words, an author got a green validate followed by a vouch that refused —
+ * two commands contradicting each other about one document, with nothing in the
+ * green run hinting at it.
+ */
+export async function expandSourceFiles(repoDir, sources, label) {
+    const { found, skipped } = await collectSources(repoDir, sources);
+    const files = found.map((f) => f.rel);
+    return { files, skipped, ...(files.length === 0 ? { empty: emptySourcesMessage(label, sources) } : {}) };
+}
+/**
+ * The one sentence for "these paths exist and cover no file". One definition,
+ * because `vouch`'s refusal and `validate`'s finding are the same diagnosis
+ * about the same document and must not drift into two descriptions of it — the
+ * author fixes it once, in the sources list.
+ */
+export function emptySourcesMessage(label, sources) {
+    return `${label}: the sources listed match no files — ${sources.join(", ")}. A digest over nothing would read as current forever.`;
+}
+/* ------------------------------------------------------------------ */
+/* The per-file index: what `sources.stale` names                      */
+/* ------------------------------------------------------------------ */
+/**
+ * How many files `sources_files` will list one by one before it gives up and
+ * records only how many there were.
+ *
+ * A readability budget, not a correctness one: the index lives in the header of
+ * a document a person reads, and a `sources: [src/]` over a legacy service can
+ * expand to thousands of files. Past the limit `sources.stale` falls back to
+ * repeating the `sources` entries — worse advice, but a spec.md whose
+ * frontmatter is longer than its requirements is worse still.
+ */
+const SOURCE_INDEX_LIMIT = 100;
+/**
+ * The `sources_files` value `loam vouch` stamps beside `sources_digest`: one
+ * `<sha>  <path>` line per file (sha256sum's layout, and its column order, so
+ * the path may contain spaces), or just the file count once the list would run
+ * past SOURCE_INDEX_LIMIT.
+ *
+ * Its whole purpose is the next `sources.stale`. `sources_digest` can say THAT
+ * the code moved and never which part of it; with the index a reader gets the
+ * added, removed and changed paths, which is the difference between re-reading
+ * one file and re-reading a directory. It is also a readable git diff in the
+ * docs repo: the vouch commit shows exactly which source files the stamp now
+ * stands over.
+ */
+export function encodeSourceIndex(index) {
+    if (index.length > SOURCE_INDEX_LIMIT)
+        return String(index.length);
+    return index.map((e) => `${e.sha}  ${e.path}`).join("\n");
+}
+/**
+ * Read back what `encodeSourceIndex` wrote. Anything unrecognised reads as "the
+ * stamp said nothing" rather than as an empty index: a header nobody can parse
+ * must not turn into a `sources.stale` claiming every file was deleted.
+ */
+export function decodeSourceIndex(stamped) {
+    if (stamped === undefined)
+        return { entries: null };
+    const text = stamped.trim();
+    if (text.length === 0)
+        return { entries: null };
+    if (/^\d+$/.test(text))
+        return { entries: null, count: Number(text) };
+    const entries = new Map();
+    for (const line of text.split("\n")) {
+        const match = /^([0-9a-f]{16})\s+(.+)$/.exec(line.trim());
+        if (match === null)
+            return { entries: null };
+        entries.set(match[2], match[1]);
+    }
+    return { entries, count: entries.size };
+}
+/**
+ * The structured half of the answer: which paths were added, changed and
+ * removed between a stamp and the current index — or null when the stamp
+ * cannot say. Null covers two distinct silences on purpose, both of which mean
+ * "do not print a path list": a stamp that recorded no per-file entries (the
+ * count-only >100-file fallback, or a header nobody could parse), and an index
+ * that AGREES with the current tree while the caller's digest comparison said
+ * stale — a contradiction (a truncation collision, or a hand-edited header),
+ * where an empty delta under a staleness warning would read as proof that
+ * nothing moved. Callers that need the counts anyway read `stamp.count`.
+ *
+ * Extracted from `movedSources` (which now consumes it) when the re-vouch
+ * reading pack became a second reader: the pack wants the three lists as
+ * data, the stale finding wants them annotated in one column.
+ */
+export function sourceDelta(stamp, now) {
+    if (stamp.entries === null)
+        return null;
+    const added = [];
+    const changed = [];
+    const current = new Map(now.map((e) => [e.path, e.sha]));
+    for (const [path, sha] of current) {
+        const before = stamp.entries.get(path);
+        if (before === undefined)
+            added.push(path);
+        else if (before !== sha)
+            changed.push(path);
+    }
+    const removed = [...stamp.entries.keys()].filter((path) => !current.has(path));
+    if (added.length + changed.length + removed.length === 0)
+        return null;
+    added.sort();
+    changed.sort();
+    removed.sort();
+    return { added, changed, removed };
+}
+/**
+ * added / removed / changed, from the stamped index against the current one.
+ * The annotation travels IN the detail line rather than in three separate
+ * lists: a reader scanning a stale report wants the paths in one column, and
+ * "which of the three" is one word wide.
+ */
+export function movedSources(stamp, now) {
+    const delta = sourceDelta(stamp, now);
+    if (delta === null) {
+        // A recorded index that came back null is the contradiction case — say
+        // nothing rather than print an empty list under a staleness warning. With
+        // no index at all, the count alone still beats silence: "12 files then,
+        // 14 now" tells a reader to go looking for two new files.
+        return {
+            summary: stamp.entries !== null || stamp.count === undefined
+                ? ""
+                : ` (${stamp.count} file(s) then, ${now.length} now)`,
+        };
+    }
+    const paths = [
+        ...delta.added.map((path) => `added    ${path}`),
+        ...delta.changed.map((path) => `changed  ${path}`),
+        ...delta.removed.map((path) => `removed  ${path}`),
+    ].sort();
+    return { paths, summary: ` (${paths.length} path(s) moved)` };
+}
+//# sourceMappingURL=stamp.js.map

package/dist/core/provenance/stamp.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"stamp.js","sourceRoot":"","sources":["../../../src/core/provenance/stamp.ts"],"names":[],"mappings":"AAAA;;;;;;;GAOG;AACH,OAAO,EAAE,UAAU,EAAE,MAAM,aAAa,CAAC;AACzC,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AAC5C,OAAO,EAAE,cAAc,EAAsB,MAAM,WAAW,CAAC;AAE/D;;;;;;GAMG;AACH,MAAM,UAAU,KAAK,CAAC,GAAS;IAC7B,MAAM,GAAG,GAAG,CAAC,CAAS,EAAU,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC,CAAC;IAC9D,OAAO,GAAG,GAAG,CAAC,WAAW,EAAE,IAAI,GAAG,CAAC,GAAG,CAAC,QAAQ,EAAE,GAAG,CAAC,CAAC,IAAI,GAAG,CAAC,GAAG,CAAC,OAAO,EAAE,CAAC,EAAE,CAAC;AACjF,CAAC;AAED,wEAAwE;AACxE,yEAAyE;AACzE,wEAAwE;AAExE,2DAA2D;AAC3D,MAAM,aAAa,GAAG,EAAE,CAAC;AAEzB,4EAA4E;AAC5E,MAAM,iBAAiB,GAAG,EAAE,CAAC;AAkC7B;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,MAAM,CAAC,KAAK,UAAU,aAAa,CAAC,OAAe,EAAE,OAAiB;IACpE,MAAM,EAAE,KAAK,EAAE,OAAO,EAAE,GAAG,MAAM,cAAc,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;IAClE,MAAM,KAAK,GAAG,UAAU,CAAC,QAAQ,CAAC,CAAC;IACnC,MAAM,KAAK,GAAuB,EAAE,CAAC;IACrC,4EAA4E;IAC5E,6EAA6E;IAC7E,4EAA4E;IAC5E,6EAA6E;IAC7E,0EAA0E;IAC1E,yEAAyE;IACzE,8EAA8E;IAC9E,oEAAoE;IACpE,KAAK,IAAI,KAAK,GAAG,CAAC,EAAE,KAAK,GAAG,KAAK,CAAC,MAAM,EAAE,KAAK,IAAI,iBAAiB,EAAE,CAAC;QACrE,MAAM,MAAM,GAAG,MAAM,OAAO,CAAC,GAAG,CAC9B,KAAK,CAAC,KAAK,CAAC,KAAK,EAAE,KAAK,GAAG,iBAAiB,CAAC,CAAC,GAAG,CAAC,KAAK,EAAE,IAAI,EAAE,EAAE,CAAC,CAAC;YACjE,GAAG,EAAE,IAAI,CAAC,GAAG;YACb,OAAO,EAAE,UAAU,CAAC,QAAQ,CAAC,CAAC,MAAM,CAAC,MAAM,QAAQ,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC;SAC7E,CAAC,CAAC,CACJ,CAAC;QACF,KAAK,MAAM,EAAE,GAAG,EAAE,OAAO,EAAE,IAAI,MAAM,EAAE,CAAC;YACtC,KAAK,CAAC,MAAM,CAAC,GAAG,GAAG,KAAK,OAAO,IAAI,CAAC,CAAC;YACrC,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,EAAE,GAAG,EAAE,GAAG,EAAE,OAAO,CAAC,KAAK,CAAC,CAAC,EAAE,aAAa,CAAC,EAAE,CAAC,CAAC;QAClE,CAAC;IACH,CAAC;IACD,uEAAuE;IACvE,4EAA4E;IAC5E,6EAA6E;IAC7E,OAAO,EAAE,MAAM,EAAE,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,EAAE,aAAa,CAAC,EAAE,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,CAAC,EAAE,KAAK,EAAE,OAAO,EAAE,CAAC;AAClH,CAAC;AAED;;;;;;;;;;;GAWG;AACH,MAAM,CAAC,KAAK,UAAU,iBAAiB,CACrC,OAAe,EACf,OAAiB,EACjB,KAAa;IAEb,MAAM,EAAE,KAAK,EAAE,OAAO,EAAE,GAAG,MAAM,cAAc,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;IAClE,MAAM,KAAK,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,GAAG,CAAC,CAAC;IACtC,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE,GAAG,CAAC,KAAK,CAAC,MAAM,KAAK,CAAC,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,mBAAmB,CAAC,KAAK,EAAE,OAAO,CAAC,EAAE,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,CAAC;AAC3G,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,mBAAmB,CAAC,KAAa,EAAE,OAAiB;IAClE,OAAO,GAAG,KAAK,yCAAyC,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,wDAAwD,CAAC;AACrI,CAAC;AAED,wEAAwE;AACxE,yEAAyE;AACzE,wEAAwE;AAExE;;;;;;;;;GASG;AACH,MAAM,kBAAkB,GAAG,GAAG,CAAC;AAE/B;;;;;;;;;;;;GAYG;AACH,MAAM,UAAU,iBAAiB,CAAC,KAAyB;IACzD,IAAI,KAAK,CAAC,MAAM,GAAG,kBAAkB;QAAE,OAAO,MAAM,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;IACnE,OAAO,KAAK,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,GAAG,CAAC,CAAC,GAAG,KAAK,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;AAC5D,CAAC;AAUD;;;;GAIG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAA2B;IAC3D,IAAI,OAAO,KAAK,SAAS;QAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IACpD,MAAM,IAAI,GAAG,OAAO,CAAC,IAAI,EAAE,CAAC;IAC5B,IAAI,IAAI,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;IAChD,IAAI,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC;QAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,KAAK,EAAE,MAAM,CAAC,IAAI,CAAC,EAAE,CAAC;IAEtE,MAAM,OAAO,GAAG,IAAI,GAAG,EAAkB,CAAC;IAC1C,KAAK,MAAM,IAAI,IAAI,IAAI,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,CAAC;QACpC,MAAM,KAAK,GAAG,yBAAyB,CAAC,IAAI,CAAC,IAAI,CAAC,IAAI,EAAE,CAAC,CAAC;QAC1D,IAAI,KAAK,KAAK,IAAI;YAAE,OAAO,EAAE,OAAO,EAAE,IAAI,EAAE,CAAC;QAC7C,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,CAAC,CAAE,EAAE,KAAK,CAAC,CAAC,CAAE,CAAC,CAAC;IACpC,CAAC;IACD,OAAO,EAAE,OAAO,EAAE,KAAK,EAAE,OAAO,CAAC,IAAI,EAAE,CAAC;AAC1C,CAAC;AAUD;;;;;;;;;;;;;;GAcG;AACH,MAAM,UAAU,WAAW,CACzB,KAAmB,EACnB,GAAuB;IAEvB,IAAI,KAAK,CAAC,OAAO,KAAK,IAAI;QAAE,OAAO,IAAI,CAAC;IACxC,MAAM,KAAK,GAAa,EAAE,CAAC;IAC3B,MAAM,OAAO,GAAa,EAAE,CAAC;IAC7B,MAAM,OAAO,GAAG,IAAI,GAAG,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC;IACzD,KAAK,MAAM,CAAC,IAAI,EAAE,GAAG,CAAC,IAAI,OAAO,EAAE,CAAC;QAClC,MAAM,MAAM,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC;QACvC,IAAI,MAAM,KAAK,SAAS;YAAE,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;aACtC,IAAI,MAAM,KAAK,GAAG;YAAE,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;IAC9C,CAAC;IACD,MAAM,OAAO,GAAG,CAAC,GAAG,KAAK,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,MAAM,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,IAAI,CAAC,CAAC,CAAC;IAC/E,IAAI,KAAK,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,GAAG,OAAO,CAAC,MAAM,KAAK,CAAC;QAAE,OAAO,IAAI,CAAC;IACtE,KAAK,CAAC,IAAI,EAAE,CAAC;IACb,OAAO,CAAC,IAAI,EAAE,CAAC;IACf,OAAO,CAAC,IAAI,EAAE,CAAC;IACf,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE,OAAO,EAAE,CAAC;AACrC,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,YAAY,CAAC,KAAmB,EAAE,GAAuB;IACvE,MAAM,KAAK,GAAG,WAAW,CAAC,KAAK,EAAE,GAAG,CAAC,CAAC;IACtC,IAAI,KAAK,KAAK,IAAI,EAAE,CAAC;QACnB,uEAAuE;QACvE,0EAA0E;QAC1E,wEAAwE;QACxE,0DAA0D;QAC1D,OAAO;YACL,OAAO,EACL,KAAK,CAAC,OAAO,KAAK,IAAI,IAAI,KAAK,CAAC,KAAK,KAAK,SAAS;gBACjD,CAAC,CAAC,EAAE;gBACJ,CAAC,CAAC,KAAK,KAAK,CAAC,KAAK,kBAAkB,GAAG,CAAC,MAAM,OAAO;SAC1D,CAAC;IACJ,CAAC;IACD,MAAM,KAAK,GAAG;QACZ,GAAG,KAAK,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,IAAI,EAAE,CAAC;QAChD,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,IAAI,EAAE,CAAC;QAClD,GAAG,KAAK,CAAC,OAAO,CAAC,GAAG,CAAC,CAAC,IAAI,EAAE,EAAE,CAAC,YAAY,IAAI,EAAE,CAAC;KACnD,CAAC,IAAI,EAAE,CAAC;IACT,OAAO,EAAE,KAAK,EAAE,OAAO,EAAE,KAAK,KAAK,CAAC,MAAM,iBAAiB,EAAE,CAAC;AAChE,CAAC","sourcesContent":["/**\n * The stamp vocabulary: what `loam vouch` computes for a frontmatter stamp and\n * what a later check reads back out of one — `sources_digest` and the per-file\n * `sources_files` index beside it, and the local calendar day `last_verified`\n * carries. The document's own `content_digest` lived here too; it moved to\n * core/document/frontmatter.ts, beside the body cut it hashes, when the fleet\n * enumeration needed it (`contentDigest`'s docblock says why).\n */\nimport { createHash } from \"node:crypto\";\nimport { readFile } from \"node:fs/promises\";\nimport { collectSources, type SkippedSource } from \"./walk.js\";\n\n/**\n * The local calendar day, for the two commands that date a stamp. A vouch or a\n * verification record is a person saying \"today I read this\", so it is their\n * date, not UTC's — `toISOString` files an evening in the Americas under\n * tomorrow. It sits here because this module already owns the stamp vocabulary\n * the date is written into.\n */\nexport function today(now: Date): string {\n  const pad = (n: number): string => String(n).padStart(2, \"0\");\n  return `${now.getFullYear()}-${pad(now.getMonth() + 1)}-${pad(now.getDate())}`;\n}\n\n/* ------------------------------------------------------------------ */\n/* The sources digest                                                  */\n/* ------------------------------------------------------------------ */\n\n/** How much of the sha256 is written into the document. */\nconst DIGEST_LENGTH = 16;\n\n/** How many source files the digest reads at once — see `sourcesDigest`. */\nconst DIGEST_READ_BATCH = 32;\n\nexport interface SourcesDigest extends SourcesExpansion {\n  /** The stamp that goes in `sources_digest`. */\n  digest: string;\n  /**\n   * Per file, its own short sha — the record `sources_files` carries so a later\n   * `sources.stale` can name what moved instead of guessing. Same order as `files`.\n   */\n  index: SourceIndexEntry[];\n}\n\n/** One file's share of the stamp: what it is called, and what it said. */\nexport interface SourceIndexEntry {\n  /** Repo-relative, `/`-separated. */\n  path: string;\n  /** sha256 of its bytes, truncated the way `sources_digest` is. */\n  sha: string;\n}\n\n/** What a `sources` list covers, and what it does not. */\nexport interface SourcesExpansion {\n  /** Repo-relative paths, sorted — exactly what a digest would hash. */\n  files: string[];\n  /** Everything the walk refused to follow. Empty in the ordinary case. */\n  skipped: SkippedSource[];\n  /**\n   * Set only when `files` is empty: the one sentence that says the list covers\n   * nothing. `loam vouch` refuses with it, `loam validate` grades it — see\n   * emptySourcesMessage for why they must be the same words.\n   */\n  empty?: string;\n}\n\n/**\n * Digest the CONTENT of the files `sources` names. The recipe is spelled out\n * because the value is written into documents, and anything reading them has to\n * be able to reproduce it:\n *\n *   1. expand each entry to repo-relative file paths (`/` separators) — a file\n *      is itself, a directory everything beneath it, minus what the repository\n *      does not consider its own (see collectSources) — sorted and de-duplicated;\n *   2. per file, `sha256(bytes)`;\n *   3. feed `<path>\\0<hex>\\n` for each file, in that order, into an outer\n *      sha256;\n *   4. keep the first 16 hex characters.\n *\n * Content, not mtime: git does not preserve modification times, so after a\n * fresh clone every file looks changed and the check would be false positives\n * end to end. Bytes survive the clone. Hashing the path alongside the content\n * means a rename registers, which is what a reader of the doc would want.\n *\n * 64 bits is a change detector, not a seal — it answers \"did this move?\", and\n * an adversary who wants a collision can have one.\n */\nexport async function sourcesDigest(repoDir: string, sources: string[]): Promise<SourcesDigest> {\n  const { found, skipped } = await collectSources(repoDir, sources);\n  const outer = createHash(\"sha256\");\n  const index: SourceIndexEntry[] = [];\n  // Read in fixed batches, hash in array order. `found` is already sorted and\n  // every file's own sha depends on nothing but its own bytes, so the sequence\n  // fed to `outer` — and therefore the digest — is unchanged by construction;\n  // only the waiting overlaps. A `sources: [src/]` over a real service repo is\n  // thousands of files, and one-await-per-file spent nearly all of its time\n  // idle. The batch is bounded rather than a single `Promise.all` over the\n  // whole list because an unbounded fan-out over a source tree runs the process\n  // out of file descriptors, which would turn a stamp into an EMFILE.\n  for (let start = 0; start < found.length; start += DIGEST_READ_BATCH) {\n    const hashed = await Promise.all(\n      found.slice(start, start + DIGEST_READ_BATCH).map(async (file) => ({\n        rel: file.rel,\n        content: createHash(\"sha256\").update(await readFile(file.abs)).digest(\"hex\"),\n      })),\n    );\n    for (const { rel, content } of hashed) {\n      outer.update(`${rel}\\0${content}\\n`);\n      index.push({ path: rel, sha: content.slice(0, DIGEST_LENGTH) });\n    }\n  }\n  // `empty` is deliberately not set here: it is a sentence about a named\n  // document, and a digest has no name to put in it. Callers that need it say\n  // emptySourcesMessage(label, sources) — the same words, from one definition.\n  return { digest: outer.digest(\"hex\").slice(0, DIGEST_LENGTH), files: index.map((e) => e.path), index, skipped };\n}\n\n/**\n * What a `sources` list actually covers — the digest's file set without the\n * hashing.\n *\n * Exported because two commands must give the SAME answer to \"does this list\n * cover anything?\". `loam vouch` refuses to stamp an expansion that covers no\n * files (a digest over nothing never changes, so the stamp would read as\n * current forever), and until `loam validate` could say the same thing in the\n * same words, an author got a green validate followed by a vouch that refused —\n * two commands contradicting each other about one document, with nothing in the\n * green run hinting at it.\n */\nexport async function expandSourceFiles(\n  repoDir: string,\n  sources: string[],\n  label: string,\n): Promise<SourcesExpansion> {\n  const { found, skipped } = await collectSources(repoDir, sources);\n  const files = found.map((f) => f.rel);\n  return { files, skipped, ...(files.length === 0 ? { empty: emptySourcesMessage(label, sources) } : {}) };\n}\n\n/**\n * The one sentence for \"these paths exist and cover no file\". One definition,\n * because `vouch`'s refusal and `validate`'s finding are the same diagnosis\n * about the same document and must not drift into two descriptions of it — the\n * author fixes it once, in the sources list.\n */\nexport function emptySourcesMessage(label: string, sources: string[]): string {\n  return `${label}: the sources listed match no files — ${sources.join(\", \")}. A digest over nothing would read as current forever.`;\n}\n\n/* ------------------------------------------------------------------ */\n/* The per-file index: what `sources.stale` names                      */\n/* ------------------------------------------------------------------ */\n\n/**\n * How many files `sources_files` will list one by one before it gives up and\n * records only how many there were.\n *\n * A readability budget, not a correctness one: the index lives in the header of\n * a document a person reads, and a `sources: [src/]` over a legacy service can\n * expand to thousands of files. Past the limit `sources.stale` falls back to\n * repeating the `sources` entries — worse advice, but a spec.md whose\n * frontmatter is longer than its requirements is worse still.\n */\nconst SOURCE_INDEX_LIMIT = 100;\n\n/**\n * The `sources_files` value `loam vouch` stamps beside `sources_digest`: one\n * `<sha>  <path>` line per file (sha256sum's layout, and its column order, so\n * the path may contain spaces), or just the file count once the list would run\n * past SOURCE_INDEX_LIMIT.\n *\n * Its whole purpose is the next `sources.stale`. `sources_digest` can say THAT\n * the code moved and never which part of it; with the index a reader gets the\n * added, removed and changed paths, which is the difference between re-reading\n * one file and re-reading a directory. It is also a readable git diff in the\n * docs repo: the vouch commit shows exactly which source files the stamp now\n * stands over.\n */\nexport function encodeSourceIndex(index: SourceIndexEntry[]): string {\n  if (index.length > SOURCE_INDEX_LIMIT) return String(index.length);\n  return index.map((e) => `${e.sha}  ${e.path}`).join(\"\\n\");\n}\n\n/** What a stamped `sources_files` says: the per-file shas, the count, or nothing. */\ninterface StampedIndex {\n  /** path -> sha, or null when the stamp did not record them. */\n  entries: Map<string, string> | null;\n  /** How many files the stamp covered, when that is knowable. */\n  count?: number;\n}\n\n/**\n * Read back what `encodeSourceIndex` wrote. Anything unrecognised reads as \"the\n * stamp said nothing\" rather than as an empty index: a header nobody can parse\n * must not turn into a `sources.stale` claiming every file was deleted.\n */\nexport function decodeSourceIndex(stamped: string | undefined): StampedIndex {\n  if (stamped === undefined) return { entries: null };\n  const text = stamped.trim();\n  if (text.length === 0) return { entries: null };\n  if (/^\\d+$/.test(text)) return { entries: null, count: Number(text) };\n\n  const entries = new Map<string, string>();\n  for (const line of text.split(\"\\n\")) {\n    const match = /^([0-9a-f]{16})\\s+(.+)$/.exec(line.trim());\n    if (match === null) return { entries: null };\n    entries.set(match[2]!, match[1]!);\n  }\n  return { entries, count: entries.size };\n}\n\n/** The paths that moved between a stamp and now, with a phrase for the message. */\ninterface MovedSources {\n  /** Annotated paths, or undefined when the stamp did not record enough to say. */\n  paths?: string[];\n  /** Appended to the `sources.stale` message; empty when there is nothing to add. */\n  summary: string;\n}\n\n/**\n * The structured half of the answer: which paths were added, changed and\n * removed between a stamp and the current index — or null when the stamp\n * cannot say. Null covers two distinct silences on purpose, both of which mean\n * \"do not print a path list\": a stamp that recorded no per-file entries (the\n * count-only >100-file fallback, or a header nobody could parse), and an index\n * that AGREES with the current tree while the caller's digest comparison said\n * stale — a contradiction (a truncation collision, or a hand-edited header),\n * where an empty delta under a staleness warning would read as proof that\n * nothing moved. Callers that need the counts anyway read `stamp.count`.\n *\n * Extracted from `movedSources` (which now consumes it) when the re-vouch\n * reading pack became a second reader: the pack wants the three lists as\n * data, the stale finding wants them annotated in one column.\n */\nexport function sourceDelta(\n  stamp: StampedIndex,\n  now: SourceIndexEntry[],\n): { added: string[]; changed: string[]; removed: string[] } | null {\n  if (stamp.entries === null) return null;\n  const added: string[] = [];\n  const changed: string[] = [];\n  const current = new Map(now.map((e) => [e.path, e.sha]));\n  for (const [path, sha] of current) {\n    const before = stamp.entries.get(path);\n    if (before === undefined) added.push(path);\n    else if (before !== sha) changed.push(path);\n  }\n  const removed = [...stamp.entries.keys()].filter((path) => !current.has(path));\n  if (added.length + changed.length + removed.length === 0) return null;\n  added.sort();\n  changed.sort();\n  removed.sort();\n  return { added, changed, removed };\n}\n\n/**\n * added / removed / changed, from the stamped index against the current one.\n * The annotation travels IN the detail line rather than in three separate\n * lists: a reader scanning a stale report wants the paths in one column, and\n * \"which of the three\" is one word wide.\n */\nexport function movedSources(stamp: StampedIndex, now: SourceIndexEntry[]): MovedSources {\n  const delta = sourceDelta(stamp, now);\n  if (delta === null) {\n    // A recorded index that came back null is the contradiction case — say\n    // nothing rather than print an empty list under a staleness warning. With\n    // no index at all, the count alone still beats silence: \"12 files then,\n    // 14 now\" tells a reader to go looking for two new files.\n    return {\n      summary:\n        stamp.entries !== null || stamp.count === undefined\n          ? \"\"\n          : ` (${stamp.count} file(s) then, ${now.length} now)`,\n    };\n  }\n  const paths = [\n    ...delta.added.map((path) => `added    ${path}`),\n    ...delta.changed.map((path) => `changed  ${path}`),\n    ...delta.removed.map((path) => `removed  ${path}`),\n  ].sort();\n  return { paths, summary: ` (${paths.length} path(s) moved)` };\n}\n"]}

package/dist/core/provenance/walk.js

@@ -0,0 +1,157 @@
+/**
+ * The source walk: expanding a `sources` list to the concrete files it covers
+ * — found and skipped. What the walk refuses to follow is reported, never
+ * dropped, because the digest downstream cannot go stale over bytes it never
+ * hashed; the skipped list is how that hole stays visible.
+ */
+import { readdir, realpath, stat } from "node:fs/promises";
+import { join, relative, resolve, sep } from "node:path";
+import { isPathInside, resolveInside } from "../kernel/path-safety.js";
+import { gitIgnoredPaths } from "./git.js";
+/**
+ * The files a `sources` list names: a file is itself, a directory is everything
+ * beneath it. Nothing else — glob patterns used to be matched here, by a
+ * hand-rolled dialect that silently differed from the gitignore/minimatch
+ * conventions authors assume, so a pattern quietly digested a DIFFERENT file
+ * set than its author intended. Pattern-looking entries are now refused loudly
+ * upstream (see patternSources); this function only ever sees literal paths.
+ *
+ * Three exclusions, all part of the digest recipe's contract:
+ *
+ *  - dot-entries are skipped while walking — `.git` is not what the doc was
+ *    written from — though a path naming one outright is still honoured;
+ *  - `node_modules` is skipped the same way, unconditionally: it is the one
+ *    build input every ecosystem agrees is not source, and it is the one that
+ *    makes the walk take minutes;
+ *  - anything the repository's own `.gitignore` covers is dropped, when the
+ *    repo is a git checkout and git is on the PATH (see gitIgnoredPaths).
+ *
+ * The third is the one worth stating a reason for. `sources_digest` answers
+ * "did the code move since a person read it", and BUILD OUTPUT moves on every
+ * CI run without anybody touching the code — a `sources: [src/]` over a repo
+ * that compiles into `src/generated/` went stale on a schedule, so the warning
+ * that means "re-read this" arrived when nothing had been written. A signal
+ * that fires every night is a signal people learn to close. Git's answer is
+ * used rather than a list of our own because it is the answer the repository
+ * already gives every other tool, and it is the author's to change.
+ *
+ * The fallback is deliberate: no git, no checkout, a git that errors — hash
+ * everything. Missing an exclusion costs noise; inventing one costs a file the
+ * stamp silently stops watching, and this walk exists to have nothing in that
+ * category.
+ */
+export async function collectSources(repoDir, sources) {
+    const found = new Map();
+    const skipped = new Map();
+    const relOf = (abs) => relative(repoDir, abs).split(sep).join("/");
+    // realpath once, up front: every containment question below is asked against
+    // the resolved root, so a repo reached through a symlink (a worktree under
+    // /var -> /private/var on macOS, say) does not make its own files look external.
+    const repoReal = await realpath(repoDir).catch(() => resolve(repoDir));
+    // The realpaths of the directories currently being walked — the cycle guard.
+    // A STACK rather than a set of everything ever entered, because the two rules
+    // answer different questions: a directory reachable by two spellings (a
+    // `current -> versions/3` link beside the real tree) is content under both
+    // names and belongs in the digest twice, while a directory reachable from
+    // INSIDE itself is a loop. A global visited-set conflates them, and then which
+    // spelling survives depends on the order readdir happened to return — a
+    // digest that differs between two machines holding identical bytes.
+    const walking = [];
+    const walk = async (dir) => {
+        const entries = await readdir(dir, { withFileTypes: true }).catch(() => []);
+        // Sorted, so a tree walks the same way on every filesystem.
+        entries.sort((a, b) => (a.name < b.name ? -1 : a.name > b.name ? 1 : 0));
+        for (const entry of entries) {
+            if (entry.name.startsWith(".") || entry.name === "node_modules")
+                continue;
+            const child = join(dir, entry.name);
+            if (entry.isDirectory())
+                await enterDir(child);
+            else if (entry.isFile())
+                found.set(relOf(child), child);
+            // readdir reports symlinks by lstat, so isDirectory() and isFile() are
+            // BOTH false for one: before this branch existed a symlinked file, and a
+            // whole subtree behind a symlinked directory, fell between the two arms
+            // and left the digest without a word said. That is the failure this
+            // module is here to prevent, arriving through its own walk.
+            else if (entry.isSymbolicLink())
+                await follow(child);
+        }
+    };
+    const enterDir = async (dir) => {
+        const real = await realpath(dir).catch(() => null);
+        if (real === null || walking.includes(real))
+            return;
+        walking.push(real);
+        try {
+            await walk(dir);
+        }
+        finally {
+            walking.pop();
+        }
+    };
+    const follow = async (link) => {
+        const real = await realpath(link).catch(() => null);
+        if (real === null) {
+            skipped.set(relOf(link), "a symlink that does not resolve");
+            return;
+        }
+        // Outside the repo the file is not this service's to vouch for, and the
+        // same rule already refuses it as a top-level `sources` entry
+        // (sources.path-outside). Reported rather than dropped: an author who
+        // vendored a sibling repo through a symlink needs to know the stamp stops
+        // at the link.
+        if (!isPathInside(repoReal, real)) {
+            skipped.set(relOf(link), "a symlink whose target is outside this repository");
+            return;
+        }
+        const info = await stat(link).catch(() => null);
+        if (info === null) {
+            skipped.set(relOf(link), "a symlink that does not resolve");
+            return;
+        }
+        // A link's own spelling is what the doc's author wrote, so that is the path
+        // that goes into the digest — the content comes from the target either way.
+        if (info.isFile())
+            found.set(relOf(link), link);
+        // A link that closes a loop is not reported: everything behind it is already
+        // in the digest under the spelling it was reached by, so nothing has stopped
+        // being watched. Only unhashed CONTENT is worth a warning.
+        else if (info.isDirectory())
+            await enterDir(link);
+        else
+            skipped.set(relOf(link), "a symlink to neither a file nor a directory");
+    };
+    for (const source of sources) {
+        const cleaned = source.trim();
+        if (cleaned.length === 0)
+            continue;
+        const root = resolveInside(repoDir, cleaned, `source '${source}'`);
+        const info = await stat(root).catch(() => null);
+        if (info === null)
+            continue;
+        if (info.isFile())
+            found.set(relOf(root), root);
+        else if (info.isDirectory())
+            await enterDir(root);
+    }
+    // One git invocation for the whole expansion, not one per file: the skipped
+    // list rides along so an ignored symlink does not warn about a path the
+    // repository already says is not its own.
+    const ignored = await gitIgnoredPaths(repoDir, [...found.keys(), ...skipped.keys()]);
+    for (const rel of ignored) {
+        found.delete(rel);
+        skipped.delete(rel);
+    }
+    return {
+        // Plain codepoint order, not locale order: the digest has to be the same
+        // everywhere it is computed.
+        found: [...found.entries()]
+            .map(([rel, abs]) => ({ rel, abs }))
+            .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0)),
+        skipped: [...skipped.entries()]
+            .map(([path, reason]) => ({ path, reason }))
+            .sort((a, b) => (a.path < b.path ? -1 : a.path > b.path ? 1 : 0)),
+    };
+}
+//# sourceMappingURL=walk.js.map

package/dist/core/provenance/walk.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"walk.js","sourceRoot":"","sources":["../../../src/core/provenance/walk.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AACH,OAAO,EAAE,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,MAAM,kBAAkB,CAAC;AAC3D,OAAO,EAAE,IAAI,EAAE,QAAQ,EAAE,OAAO,EAAE,GAAG,EAAE,MAAM,WAAW,CAAC;AACzD,OAAO,EAAE,YAAY,EAAE,aAAa,EAAE,MAAM,0BAA0B,CAAC;AACvE,OAAO,EAAE,eAAe,EAAE,MAAM,UAAU,CAAC;AAgB3C;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AACH,MAAM,CAAC,KAAK,UAAU,cAAc,CAClC,OAAe,EACf,OAAiB;IAEjB,MAAM,KAAK,GAAG,IAAI,GAAG,EAAkB,CAAC;IACxC,MAAM,OAAO,GAAG,IAAI,GAAG,EAAkB,CAAC;IAC1C,MAAM,KAAK,GAAG,CAAC,GAAW,EAAU,EAAE,CAAC,QAAQ,CAAC,OAAO,EAAE,GAAG,CAAC,CAAC,KAAK,CAAC,GAAG,CAAC,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;IAEnF,6EAA6E;IAC7E,2EAA2E;IAC3E,iFAAiF;IACjF,MAAM,QAAQ,GAAG,MAAM,QAAQ,CAAC,OAAO,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC;IACvE,6EAA6E;IAC7E,8EAA8E;IAC9E,wEAAwE;IACxE,2EAA2E;IAC3E,0EAA0E;IAC1E,+EAA+E;IAC/E,wEAAwE;IACxE,oEAAoE;IACpE,MAAM,OAAO,GAAa,EAAE,CAAC;IAE7B,MAAM,IAAI,GAAG,KAAK,EAAE,GAAW,EAAiB,EAAE;QAChD,MAAM,OAAO,GAAG,MAAM,OAAO,CAAC,GAAG,EAAE,EAAE,aAAa,EAAE,IAAI,EAAE,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,EAAE,CAAC,CAAC;QAC5E,4DAA4D;QAC5D,OAAO,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QACzE,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;YAC5B,IAAI,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,CAAC,IAAI,KAAK,CAAC,IAAI,KAAK,cAAc;gBAAE,SAAS;YAC1E,MAAM,KAAK,GAAG,IAAI,CAAC,GAAG,EAAE,KAAK,CAAC,IAAI,CAAC,CAAC;YACpC,IAAI,KAAK,CAAC,WAAW,EAAE;gBAAE,MAAM,QAAQ,CAAC,KAAK,CAAC,CAAC;iBAC1C,IAAI,KAAK,CAAC,MAAM,EAAE;gBAAE,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,KAAK,CAAC,EAAE,KAAK,CAAC,CAAC;YACxD,uEAAuE;YACvE,yEAAyE;YACzE,wEAAwE;YACxE,oEAAoE;YACpE,4DAA4D;iBACvD,IAAI,KAAK,CAAC,cAAc,EAAE;gBAAE,MAAM,MAAM,CAAC,KAAK,CAAC,CAAC;QACvD,CAAC;IACH,CAAC,CAAC;IAEF,MAAM,QAAQ,GAAG,KAAK,EAAE,GAAW,EAAiB,EAAE;QACpD,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,GAAG,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC;QACnD,IAAI,IAAI,KAAK,IAAI,IAAI,OAAO,CAAC,QAAQ,CAAC,IAAI,CAAC;YAAE,OAAO;QACpD,OAAO,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACnB,IAAI,CAAC;YACH,MAAM,IAAI,CAAC,GAAG,CAAC,CAAC;QAClB,CAAC;gBAAS,CAAC;YACT,OAAO,CAAC,GAAG,EAAE,CAAC;QAChB,CAAC;IACH,CAAC,CAAC;IAEF,MAAM,MAAM,GAAG,KAAK,EAAE,IAAY,EAAiB,EAAE;QACnD,MAAM,IAAI,GAAG,MAAM,QAAQ,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC;QACpD,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;YAClB,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,iCAAiC,CAAC,CAAC;YAC5D,OAAO;QACT,CAAC;QACD,wEAAwE;QACxE,8DAA8D;QAC9D,sEAAsE;QACtE,0EAA0E;QAC1E,eAAe;QACf,IAAI,CAAC,YAAY,CAAC,QAAQ,EAAE,IAAI,CAAC,EAAE,CAAC;YAClC,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,mDAAmD,CAAC,CAAC;YAC9E,OAAO;QACT,CAAC;QACD,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC;QAChD,IAAI,IAAI,KAAK,IAAI,EAAE,CAAC;YAClB,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,iCAAiC,CAAC,CAAC;YAC5D,OAAO;QACT,CAAC;QACD,4EAA4E;QAC5E,4EAA4E;QAC5E,IAAI,IAAI,CAAC,MAAM,EAAE;YAAE,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,CAAC,CAAC;QAChD,6EAA6E;QAC7E,6EAA6E;QAC7E,2DAA2D;aACtD,IAAI,IAAI,CAAC,WAAW,EAAE;YAAE,MAAM,QAAQ,CAAC,IAAI,CAAC,CAAC;;YAC7C,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,6CAA6C,CAAC,CAAC;IAC/E,CAAC,CAAC;IAEF,KAAK,MAAM,MAAM,IAAI,OAAO,EAAE,CAAC;QAC7B,MAAM,OAAO,GAAG,MAAM,CAAC,IAAI,EAAE,CAAC;QAC9B,IAAI,OAAO,CAAC,MAAM,KAAK,CAAC;YAAE,SAAS;QACnC,MAAM,IAAI,GAAG,aAAa,CAAC,OAAO,EAAE,OAAO,EAAE,WAAW,MAAM,GAAG,CAAC,CAAC;QACnE,MAAM,IAAI,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,CAAC,KAAK,CAAC,GAAG,EAAE,CAAC,IAAI,CAAC,CAAC;QAChD,IAAI,IAAI,KAAK,IAAI;YAAE,SAAS;QAC5B,IAAI,IAAI,CAAC,MAAM,EAAE;YAAE,KAAK,CAAC,GAAG,CAAC,KAAK,CAAC,IAAI,CAAC,EAAE,IAAI,CAAC,CAAC;aAC3C,IAAI,IAAI,CAAC,WAAW,EAAE;YAAE,MAAM,QAAQ,CAAC,IAAI,CAAC,CAAC;IACpD,CAAC;IAED,4EAA4E;IAC5E,wEAAwE;IACxE,0CAA0C;IAC1C,MAAM,OAAO,GAAG,MAAM,eAAe,CAAC,OAAO,EAAE,CAAC,GAAG,KAAK,CAAC,IAAI,EAAE,EAAE,GAAG,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,CAAC;IACrF,KAAK,MAAM,GAAG,IAAI,OAAO,EAAE,CAAC;QAC1B,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;QAClB,OAAO,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC;IACtB,CAAC;IAED,OAAO;QACL,yEAAyE;QACzE,6BAA6B;QAC7B,KAAK,EAAE,CAAC,GAAG,KAAK,CAAC,OAAO,EAAE,CAAC;aACxB,GAAG,CAAC,CAAC,CAAC,GAAG,EAAE,GAAG,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,EAAE,GAAG,EAAE,CAAC,CAAC;aACnC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;QAC/D,OAAO,EAAE,CAAC,GAAG,OAAO,CAAC,OAAO,EAAE,CAAC;aAC5B,GAAG,CAAC,CAAC,CAAC,IAAI,EAAE,MAAM,CAAC,EAAE,EAAE,CAAC,CAAC,EAAE,IAAI,EAAE,MAAM,EAAE,CAAC,CAAC;aAC3C,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,GAAG,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;KACpE,CAAC;AACJ,CAAC","sourcesContent":["/**\n * The source walk: expanding a `sources` list to the concrete files it covers\n * — found and skipped. What the walk refuses to follow is reported, never\n * dropped, because the digest downstream cannot go stale over bytes it never\n * hashed; the skipped list is how that hole stays visible.\n */\nimport { readdir, realpath, stat } from \"node:fs/promises\";\nimport { join, relative, resolve, sep } from \"node:path\";\nimport { isPathInside, resolveInside } from \"../kernel/path-safety.js\";\nimport { gitIgnoredPaths } from \"./git.js\";\n\n/** A path the walk found but would not hash, and the reason a reader needs. */\nexport interface SkippedSource {\n  /** Repo-relative, `/`-separated. */\n  path: string;\n  /** Why it was left out — the half of the sentence after the path. */\n  reason: string;\n}\n\nexport interface SourceFile {\n  /** Repo-relative, `/`-separated — the spelling that goes into the hash. */\n  rel: string;\n  abs: string;\n}\n\n/**\n * The files a `sources` list names: a file is itself, a directory is everything\n * beneath it. Nothing else — glob patterns used to be matched here, by a\n * hand-rolled dialect that silently differed from the gitignore/minimatch\n * conventions authors assume, so a pattern quietly digested a DIFFERENT file\n * set than its author intended. Pattern-looking entries are now refused loudly\n * upstream (see patternSources); this function only ever sees literal paths.\n *\n * Three exclusions, all part of the digest recipe's contract:\n *\n *  - dot-entries are skipped while walking — `.git` is not what the doc was\n *    written from — though a path naming one outright is still honoured;\n *  - `node_modules` is skipped the same way, unconditionally: it is the one\n *    build input every ecosystem agrees is not source, and it is the one that\n *    makes the walk take minutes;\n *  - anything the repository's own `.gitignore` covers is dropped, when the\n *    repo is a git checkout and git is on the PATH (see gitIgnoredPaths).\n *\n * The third is the one worth stating a reason for. `sources_digest` answers\n * \"did the code move since a person read it\", and BUILD OUTPUT moves on every\n * CI run without anybody touching the code — a `sources: [src/]` over a repo\n * that compiles into `src/generated/` went stale on a schedule, so the warning\n * that means \"re-read this\" arrived when nothing had been written. A signal\n * that fires every night is a signal people learn to close. Git's answer is\n * used rather than a list of our own because it is the answer the repository\n * already gives every other tool, and it is the author's to change.\n *\n * The fallback is deliberate: no git, no checkout, a git that errors — hash\n * everything. Missing an exclusion costs noise; inventing one costs a file the\n * stamp silently stops watching, and this walk exists to have nothing in that\n * category.\n */\nexport async function collectSources(\n  repoDir: string,\n  sources: string[],\n): Promise<{ found: SourceFile[]; skipped: SkippedSource[] }> {\n  const found = new Map<string, string>();\n  const skipped = new Map<string, string>();\n  const relOf = (abs: string): string => relative(repoDir, abs).split(sep).join(\"/\");\n\n  // realpath once, up front: every containment question below is asked against\n  // the resolved root, so a repo reached through a symlink (a worktree under\n  // /var -> /private/var on macOS, say) does not make its own files look external.\n  const repoReal = await realpath(repoDir).catch(() => resolve(repoDir));\n  // The realpaths of the directories currently being walked — the cycle guard.\n  // A STACK rather than a set of everything ever entered, because the two rules\n  // answer different questions: a directory reachable by two spellings (a\n  // `current -> versions/3` link beside the real tree) is content under both\n  // names and belongs in the digest twice, while a directory reachable from\n  // INSIDE itself is a loop. A global visited-set conflates them, and then which\n  // spelling survives depends on the order readdir happened to return — a\n  // digest that differs between two machines holding identical bytes.\n  const walking: string[] = [];\n\n  const walk = async (dir: string): Promise<void> => {\n    const entries = await readdir(dir, { withFileTypes: true }).catch(() => []);\n    // Sorted, so a tree walks the same way on every filesystem.\n    entries.sort((a, b) => (a.name < b.name ? -1 : a.name > b.name ? 1 : 0));\n    for (const entry of entries) {\n      if (entry.name.startsWith(\".\") || entry.name === \"node_modules\") continue;\n      const child = join(dir, entry.name);\n      if (entry.isDirectory()) await enterDir(child);\n      else if (entry.isFile()) found.set(relOf(child), child);\n      // readdir reports symlinks by lstat, so isDirectory() and isFile() are\n      // BOTH false for one: before this branch existed a symlinked file, and a\n      // whole subtree behind a symlinked directory, fell between the two arms\n      // and left the digest without a word said. That is the failure this\n      // module is here to prevent, arriving through its own walk.\n      else if (entry.isSymbolicLink()) await follow(child);\n    }\n  };\n\n  const enterDir = async (dir: string): Promise<void> => {\n    const real = await realpath(dir).catch(() => null);\n    if (real === null || walking.includes(real)) return;\n    walking.push(real);\n    try {\n      await walk(dir);\n    } finally {\n      walking.pop();\n    }\n  };\n\n  const follow = async (link: string): Promise<void> => {\n    const real = await realpath(link).catch(() => null);\n    if (real === null) {\n      skipped.set(relOf(link), \"a symlink that does not resolve\");\n      return;\n    }\n    // Outside the repo the file is not this service's to vouch for, and the\n    // same rule already refuses it as a top-level `sources` entry\n    // (sources.path-outside). Reported rather than dropped: an author who\n    // vendored a sibling repo through a symlink needs to know the stamp stops\n    // at the link.\n    if (!isPathInside(repoReal, real)) {\n      skipped.set(relOf(link), \"a symlink whose target is outside this repository\");\n      return;\n    }\n    const info = await stat(link).catch(() => null);\n    if (info === null) {\n      skipped.set(relOf(link), \"a symlink that does not resolve\");\n      return;\n    }\n    // A link's own spelling is what the doc's author wrote, so that is the path\n    // that goes into the digest — the content comes from the target either way.\n    if (info.isFile()) found.set(relOf(link), link);\n    // A link that closes a loop is not reported: everything behind it is already\n    // in the digest under the spelling it was reached by, so nothing has stopped\n    // being watched. Only unhashed CONTENT is worth a warning.\n    else if (info.isDirectory()) await enterDir(link);\n    else skipped.set(relOf(link), \"a symlink to neither a file nor a directory\");\n  };\n\n  for (const source of sources) {\n    const cleaned = source.trim();\n    if (cleaned.length === 0) continue;\n    const root = resolveInside(repoDir, cleaned, `source '${source}'`);\n    const info = await stat(root).catch(() => null);\n    if (info === null) continue;\n    if (info.isFile()) found.set(relOf(root), root);\n    else if (info.isDirectory()) await enterDir(root);\n  }\n\n  // One git invocation for the whole expansion, not one per file: the skipped\n  // list rides along so an ignored symlink does not warn about a path the\n  // repository already says is not its own.\n  const ignored = await gitIgnoredPaths(repoDir, [...found.keys(), ...skipped.keys()]);\n  for (const rel of ignored) {\n    found.delete(rel);\n    skipped.delete(rel);\n  }\n\n  return {\n    // Plain codepoint order, not locale order: the digest has to be the same\n    // everywhere it is computed.\n    found: [...found.entries()]\n      .map(([rel, abs]) => ({ rel, abs }))\n      .sort((a, b) => (a.rel < b.rel ? -1 : a.rel > b.rel ? 1 : 0)),\n    skipped: [...skipped.entries()]\n      .map(([path, reason]) => ({ path, reason }))\n      .sort((a, b) => (a.path < b.path ? -1 : a.path > b.path ? 1 : 0)),\n  };\n}\n"]}