SPEC-134 is precise and costs a stamp per invocation, so it only ever covers the references someone already suspected. This is the cheap, wide signal underneath it.
Documentation already declares edges into the repository. Indexing them costs nothing per reference and needs no new syntax, no marker, and no change to a single existing page.
a97869dd4bc689efb30c581a138f14707a7d40fe65Worth stating because the milestone otherwise reads as one long chain. Phase 1 extracts edges from path= attributes, which exist today. symbol= → resolved line range is only needed for --precise, which is not in this milestone. So this item can start on day one, in parallel with WORK-597.
The temptation is a function that walks content, scores as it goes, and prints. That shape cannot answer touching, because WORK-596's query runs in the opposite direction and without the git scan at all.
Extract edges into an addressable index first and let the ranking be one reader of it. The cost is an afternoon here and a rewrite avoided later.
| Piece | State |
|---|---|
Single-pass git log --format=%at --name-only scan | packages/content/src/timestamps.ts:55 — the exact scan shape |
| Shallow-clone detection | timestamps.ts:37 — isShallowClone, already written |
| Git-root-relative path prefixing | timestamps.ts:123 — getGitRelativePrefix |
| Embedded-source edge extraction | the shared reader, read-file.ts |
A ranked report with --format json | plan status, the pattern to follow |
A shallow clone reports an empty result that is indistinguishable from a clean corpus. That is the failure this command cannot afford, because its output is already "nothing is very wrong" most of the time.
The spec is explicit: embedded-source edges are the class whose extraction is already trustworthy, so they are the right one to establish the index, the scan and the output shape against. WORK-595 is where the corpus is actually reached.
stale command is registered in packages/cli/src/commands/ — no such command exists todayrefrakt stale reports a ranked list of edges, ordered by commits to the target since the referrer last changedgit log --name-only pass over the repository, not one git log invocation per edge--full-history and no pathspec, so history simplification cannot drop commits from a file's historysnippet, file-ref and expand path= attributes--top, defaulting to 10, and each entry lists commit subjects for the target's changes--class, --min and --format json behave as specifiedrefrakt stale exits zero whatever it finds, including when every edge in the corpus is stalefetch-depth: 0site/content/docs/cli/cli-overview.md gains a row for stale — the page WORK-595 cites as its motivating instance of a stale command tableAdapt timestamps.ts's scan rather than rewriting it; the shallow-clone and git-root handling there are the parts most easily got wrong.
Pass --full-history, and do not take the existing call shape on trust. Found while measuring SPEC-136 D16's base rate, and it is the single easiest way to ship this item with quietly wrong numbers:
git log --name-only -- site/content/docs/cli/cli-overview.md → one commit, 2026-07-24 git log --full-history --name-only (no pathspec) → five commits, most recent 2026-09-10
A pathspec makes git apply history simplification. Both outputs are correct git answering different questions, but this command does arithmetic on those timestamps — the referrer's last-changed time is one half of every score — so a simplified history inflates every edge out of that file, and deflates every edge into it.
The failure is invisible: the report still renders, still ranks, still looks plausible. It is the spec's own subject turned on the spec's own instrument, which is why it gets a pinned test rather than a comment.
The commit subjects are the report's whole value. "11 commits" is a number; "generate breadcrumb and timeline positions from a declared index" is a person recognising that the page they wrote does not mention declared indexes. Do not economise on them.
This ranks. It never fails (D1). No exit code on findings, no build failure, no required check, not even an opt-in flag to make it failing.
The failure mode of getting that wrong is specific and fatal: the cheapest way to turn any edge green is to edit the referring page, so a gate would train people to make trivial documentation edits to clear it — actively destroying the signal it measures. The measure only survives if nothing depends on it being zero.
Two edge classes are rejected, not deferred: plan source= edges (99% base rate) and derived rune pages (96%). Both are in the spec's table with their measurements; do not re-add them.
--top and commit subjects), D12 (the index as a shared module)touching query this index must be able to answerpackages/content/src/timestamps.ts — the single-pass scan and isShallowClone, both adapted rather than rewrittenpackages/runes/src/lib/read-file.ts — the shared reader carrying path=Completed: 2026-09-24
Branch: claude/v0-37-0-review-vqpl41 PR: refrakt-md/refrakt#648 (batched with WORK-590)
packages/content/src/edges/ (new) — extract.ts, git-scan.ts and index.ts. The index is addressable in both directions; ranking is one reader of it.packages/cli/src/commands/stale.ts (new) — refrakt stale, registered in bin.ts with --top, --class, --min, --format.packages/content/test/edges.test.ts — 21 tests.site/content/docs/cli/cli-overview.md gains the row.touching even though WORK-596 ships it. A walk-score-print function cannot answer a query that runs target → referrer and without git. edgesTouching already exists and is tested, including for a path with no commit history at all.--full-history test needed a real divergence. My first fixture reported 2 commits both ways — it would have passed while proving nothing. A merge using -s ours (discarding the side's version) is the construction that diverges: simplified 2, scan 3. Worth knowing if this fixture is ever edited.fetch-depth: 0 before the test existed. git fetch --unshallow was needed to see any ranking at all.{% snippet %} inside a ```markdoc fence inflated this repo's index from 6 real edges to 10. Reuses WORK-588's Markdown masker rather than re-deriving fence tracking.4858 tests pass. Live run against this repository produces one genuine finding.