docs(metadata-protocol): re-anchor the remaining dead tracker and comment-id citations in test comments (stage 2 of #20595) (#21246)

Part of #20595
Clause-②: no

## What changed

Stage 2 of the `domain:engine` lane of the dead-citation sweep: the
`packages/metadata-protocol` remainder that stage 1 (PR #21233, landed
as `a7d9768ec`) left, per the claim (`5940298570`). Comment and docblock
prose only. #20595 stays open for the next stages (`objectql` next).

The population is what stage 1's contract review (`5939794396`)
escalated to this sub-stage:

- the **22 test-comment sites carrying 13 numbers that answer 404**,
which the census never reads (it defers test files);
- the **dead comment-id citation** `comment 5299845282`, on two lines;
- the **dead convenience number** `PR #14767` beside its anchor
`96326040f`.

Each is rewritten in ruling C+D's form C (record `5749154545` on
#19123): the commit in this repository's history that made the decision
the sentence describes. **23 lines in 16 files** (21 lines for the 22
numbered sites, 2 for the comment id). **Anchors: 13 numbers and the
comment id, all by commit sha; 0 by ADR or ruling record (none exists
for any of them); 0 by words alone.** Seven numbers reuse the commit
another lane already chose for them, `#14767` and the comment id keep
stage 1's commit already on their line, and five were measured here.

Every file keeps its line count (+23 / -23). No code token moves (guard
below). No citation number is added: the only numbers left on changed
lines, `#9612` and `#13331`, stood there before and answer as issues.

**No changeset, and `skip-changeset` is the declaration:** 22 of the 23
lines are in test files, which the `tsup` entry graph never reaches, and
the one `protocol.ts` line is a `//` comment inside a method body that
the build drops. Measured, not assumed (see Changeset).

## Population, before and after

**H1, re-enumerated on `origin/main` (`a7d9768ec`).** The instrument is
the gate's own code, read-only. `extractCitations` from
`scripts/check-issue-citations.mjs` runs over every tracked file in the
package (239), in both projections (whole file and comment prose),
keeping the citations `namesThisRepository` assigns to this board.
`classifyCitation` judges each one against a board read by the gate's
own `enumerateBoard`: 191 pages, frontier #21241, 19,062 records, read
20:54:38Z to 20:58:19Z on 2026-10-01. Then one REST read per number: all
13 answer **404**, and the lit controls `#5286` and `#12624` answer 200.

| reading | tree | citations | dead | src comment | test comment | test
string | changelog |
|---|---|---|---|---|---|---|---|
| before | base `a7d9768ec` | 6,092 | **141** | 0 | **22** | 63 | 56 |
| after | head `16b5bf6c1` | 6,070 | **119** | 0 | **0** | 63 | 56 |

- The before reading matches stage 1's own after reading (141: 0 / 22 /
63 / 56).
- The drop of 22 citations is exactly the rewritten numbered sites. No
live count moved: test comments resolve 1,586 and
resolve-as-pull-request 111 at both.
- **The count against the dispatch's 22 / 13: 22 sites, 13 numbers, on
21 lines.** One line carries two numbers. The lines are in **14** test
files; the review counted 15.
- Comment ids (10-digit, outside the citation grammar): 9 sites, 5
distinct ids at the base. Read one by one through the comments endpoint,
`5299845282` answers **404** and the other four answer 200: `5302933802`
(on #8833), `5528370129` (#13609), `5865708652` (#20378) and
`5875579209` (#20451). The control `5940298570` (on #20595) answers 200.
After: 7 sites, 0 dead.

## Per-number table

Every sha below resolves to exactly one commit (`git rev-parse
--disambiguate`). Each is an ancestor of the base: `git merge-base
--is-ancestor` exits 0 for all 14, on a full clone
(`--is-shallow-repository` is false). Each one's message or diff names
the number it replaces. `ADR`: `git grep` over `docs/adr` and
`scripts/adr-anchors` names none of the 13 numbers or the comment id, so
the ADR rung is empty for all of them.

| number | sites | anchor | kind | named in | provenance | what it
decided |
|---|---|---|---|---|---|---|
| `#6287` | 1 | `84c86fb45` | commit | message (squash suffix) + diff |
reused: runtime and spec lanes | `preview` / `trial` discovery folds
declared, the fold table exhaustive over `EnvironmentType` |
| `#10058` | 1 | `6f5a44976` | commit | subject (the squash of that PR)
| measured here | judge a package write against its own closure |
| `#10064` | 2 | `def0d3e63` | commit | message + diff | reused: lint
lane | key publish-gate finding paths by name, not the private snapshot
index |
| `#10420` | 1 | `05bc692e0` | commit | message + diff | measured here |
run `SysMetadataRepository` through the shared repository contract suite
(it adds this file) |
| `#10978` | 2 | `4c9780c7a` | commit | message + diff | reused: runtime
lane | ObjectQL doubles enforce `limit` after the filter and by presence
|
| `#11017` | 1 | `d806081dd` | commit | diff (it adds this file) |
measured here | render the 422 findings clause per write face |
| `#13214` | 1 | `cc837dbfe` | commit | diff | reused: rest lane |
require the resolved environment to belong to the caller at `GET
/ui/view/:object/:type` |
| `#13244` | 1 | `889ec5b42` | commit | subject (squash) | reused: rest
lane | measure identity resolution at `GET /ui/view/:object/:type` |
| `#13258` | 1 | `3d10755f0` | commit | subject (squash) | reused: rest
lane | drive the tenancy axis of `GET /ui/view/:object/:type` (it adds
the sibling harness) |
| `#14389` | 1 | `10220a7bf` | commit | diff | reused: rest lane | the
409 arm gated on the engine's envelope (name and code) |
| `#14431` | 5 | `a98b61b3e` | commit | subject (squash) | measured here
| re-measure the deleted-datasource prolongation across replicas (it
adds this file) |
| `#14767` | 1 | `96326040f` | commit | subject (squash) | stage 1's
anchor, already on the line | the dead convenience number is dropped;
the sha stays |
| `#17621` | 4 | `7e74af3df` | commit | message + diff | measured here |
execute the read-probe PostgreSQL arm on a live server (it adds the
live-postgres file) |
| `comment 5299845282` | 2 | `75e66fc8e` | commit | message | stage 1's
anchor, already beside both lines | its message records the ruling: diff
raw, then redact (Option B) |

Sentences that credit a commit with a finding, a ruling or a record were
checked against that commit's own message:

- `a98b61b3e` carries the finding all five `#14431` sites describe: the
bridge heals the registry, a read inside the residue window re-hydrates
the deleted row, and the window is bounded only when no read lands in
it.
- `7e74af3df` records the no-coverage state: the PostgreSQL arm was
pinned and run nowhere.
- `6f5a44976` records the package-closure ruling.
- `75e66fc8e` carries the Option B ruling.
- `10220a7bf` records the name-and-code gate.
- `4c9780c7a` records the after-the-filter, by-presence bound.
- `d806081dd` records the per-face decision (silence keeps the full
prose on the duplicate face) but not the ruling itself. So that site
keeps the ruling's own date and now reads 「landed as」. This is stage 1's
precedent for `#9741` and `2a29caa53`.

## Wordings to check

Most rewrites swap the tag in place, as the landed stages do: `(#N)`
becomes `(commit SHA)`, `[#N]` becomes `[commit SHA]`, and `PR #N`
becomes `commit SHA`. These say more than the tag:

- `protocol.invalid-metadata-422-face-inventory.test.ts:22`: 「The
maintainer ruling on #11017 (2026-08-22, option D) resolved it by」
became 「The maintainer ruling of 2026-08-22 (option D, landed as commit
d806081) resolved it by」.
- `protocol.ui-view-hidden-columns.test.ts:46`: 「(#13214 / PR #13258)
drives the same defect」 became 「(commit 3d10755, for the card commit
cc837db fixed) drives the same defect」. Line `:47` says 「It belongs to
that card」, and this wording keeps an antecedent for it.
- `protocol.batch-row-http-status.test.ts:546`: 「(#14389 §5)」 became
「(commit 10220a7)」. The `§5` named a section of the deleted card's
body, so it goes with the number.
- `get-meta-item-org-read-gate.test.ts:10`: 「Commit 9632604 (PR
#14767) moved」 became 「Commit 9632604 moved」. Ruling C allows a PR
number only as a convenience link, and a link that answers 404 is not
one.
- `protocol.ts:22793`: 「maintainer ruling (comment 5299845282), Option
B.」 became 「maintainer ruling (recorded in that commit's message),
Option B.」. The line above opens with `[commit 75e66fc]`.
- `protocol.diff-credential-redaction.test.ts:19`: 「(maintainer ruling,
issue comment 5299845282 — Option B)」 became 「(maintainer ruling,
recorded in commit 75e66fc's message — Option B)」.
- `migrations/seed-tenancy-backfill.live-postgres.test.ts:55` and
`:256`: 「the no-coverage state #17621 records」 became 「the no-coverage
state commit 7e74af3 records」, and 「#17621 exists because it was」
became 「commit 7e74af3 exists because it was」.
- No line was reflowed, so some lines now run past their block's wrap.
`eslint.config.mjs` declares no line-length rule, and a reflow would
move neighbouring lines and every line citation into the file.

## Sites left in this package

- **Dead tracker numbers in comment prose (src, tests and
`tsup.config.ts`): none.** Dead comment ids: none.
- **63 test-string sites** (`describe` / `it` titles and assertion
arguments) are not edited. 56 carry census-dead numbers and 7 carry
`#17621`. They are recorded on #20595, and the claim excludes string
literals.
- **`CHANGELOG.md`, 56 sites**, is release-owned and not edited.

## Mechanical guard: no code token moves

The guard compares the base blob with the working tree for all 16
touched files, with TypeScript 6.0.3, on two readings:

- **Reading 1:** the parser's leaf nodes, from a `forEachChild` walk.
Comments are trivia there, and JSDoc is never visited.
- **Reading 2:** the full token stream in parser context, from a
`getChildren` walk, with JSDoc nodes skipped.

String, template and numeric literals are compared in full on both
readings.

- **Real run at head `16b5bf6c1`:** 104,684 base tokens, **0 files with
a token change** (exit 0).
- **Controls,** each through `scripts/ablation-replace.mjs` in wrap
mode, under a shell trap that restores by absolute path from `HEAD`:
- comment (`bodies untouched` to `bodies UNTOUCHED`, `protocol.ts`): 0
files changed, exit 0;
- identifier (`const issues` to `const issuesX`,
`runtime-authoring-gate.dataset-writes.test.ts`): DIFFER on both
readings, exit 1;
- string (`'organization probe'` to `'organization probeX'`,
live-postgres test): DIFFER on both, exit 1;
- numeric (`WINDOWS_PROBED = 10` to `11`, prolongation test): DIFFER on
both, exit 1.
- **Each control landed:** anchor count went from 1 to 0 and the blob
changed. Each restore equals its `HEAD` blob (`766d14add061`,
`9c4bcfec727d`, `3457d6e431ed`, `9931944568bb`). `git diff HEAD` is
empty and the tree is clean afterwards.

## Changeset: none, with `skip-changeset` (`dist` measured)

`files[]` is `dist`, `README.md` and `CHANGELOG.md`. The `tsup` entry is
`src/index.ts`, and no test file is in its graph. For the one non-test
line (H3), the dependency closure was built first (`turbo run build
--filter='@objectstack/metadata-protocol^...'`, 12 of 12 tasks). Then
the package's own `build` ran four times under the shared verify lock,
and each leg hashed all 24 `dist` files:

| leg | state of `protocol.ts` | `dist` against leg 1 |
|---|---|---|
| 1 | head | (reference) |
| 2 | the base text put back on `:22793`, blob `5be50ab59075` (stage 1's
recorded head blob for this file) | **0 of 24 files differ**; the marker
`comment 5299845282` is in no built file |
| positive control | a docblock line on an exported member changed
(`refusal is being` to `refusal is BEING`) | `index.d.ts` and
`index.d.cts` differ; `ablation-dist-preflight` finds the marker in 2
built files |
| 3 | head again | 0 of 24 differ (deterministic);
`ablation-dist-preflight --absent` passes |

So a comment change that ships is visible to this instrument, and this
one does not ship. Restore proof: the `protocol.ts` blob equals `HEAD`
(`766d14add061`), `git diff HEAD` is empty and the tree is clean.

## Gates (head `16b5bf6c1`)

- **Derived gates:** `node scripts/pm/dispatch-gates.mjs --repo
objectstack-ai/objectstack --commands` derives 55 commands for 16 paths
against merge base `a7d9768ec`. All 55 ran, each exit captured before
any pipe, and **all 55 exit 0**. `--ran`, given `COMMAND :: exit CODE`
lines, reports 「55 derived, 55 run, 0 NOT-MEASURED, 0 UNRUN」, a derived
zero, and exits 0. A full `turbo run build` over `./packages/*` and
`./packages/*/*` ran first under the verify lock (71 of 71 tasks), so no
gate read an unbuilt workspace.
- **Named readings among them:**
- `node scripts/check-issue-citations.mjs` exits 0: 「no issue citations
added against a7d9768」.
- `pnpm check:issue-citations` exits 0: self-test, 173 cases, 9
batteries.
- `pnpm check:doc-authoring` exits 0, and the sibling-package prose-id
baseline holds with no growth.
- `pnpm check:nul-bytes` exits 0 (9,902 files). A raw control-byte scan
of the 16 changed files finds none (grep exit 1).
- **Tests and typecheck, under the verify lock:**
- `pnpm --filter @objectstack/metadata-protocol test`: 200 test files
pass and 3 skip (203); 2,973 tests pass and 19 skip. These are the same
totals as stage 1's.
- `pnpm --filter @objectstack/metadata-protocol typecheck` exits 0, and
`tsc --noEmit --listFiles` puts all 203 tracked test files and all 16
changed files in the program.
- **Lint, as a proven narrowing:** eslint runs with inline config off
(`--format json`) over the 16 touched files plus `dist/index.js` as the
control. It gives 17 results, 0 errors and 1 warning, the control's
ignore notice, and none of the 16 is reported ignored.
`eslint.config.mjs:327-328` states that no type-aware linting is
enabled, so a comment edit cannot move the verdict on an untouched file.
The repo-wide `pnpm lint` is CI's run.

## Acceptance notes

- **Base.** The branch is on `main` at `a7d9768ec`, the stage 1 squash.
`main` has moved two commits since (`1a4c7f826`, `ef96c9ede`). Neither
touches this package, `scripts/check-issue-citations.mjs` or
`scripts/pm/dispatch-gates.mjs`. No merge was taken; the merge queue
rebuilds on the merged generation.
- **File count.** The review recorded the 22 sites as 「21 lines in 15
test files」. Measured here: 21 lines in 14 files. The site and number
counts agree.
- **The same family in other packages, not edited.** Two other packages
carry similar test comments: `packages/rest` (`#14389 §5` in
`error-response-sandbox-arm-message.test.ts` and
`rest-duplicate-record-arm.test.ts`) and `packages/lint` (`(#10064)` in
`runtime-gate.object-writes.test.ts` and
`validate-object-field-refs.test.ts`). They are independent sites with
the same rot, outside this claim's file surface. Their lanes' anchors
are `10220a7bf` and `def0d3e63`.
- **The census tool was not re-run.** Its surface defers test files, so
this population is outside it by construction. The same extractor and
classifier, over the census's comment-prose projection of the package's
non-test files, read 0 dead both before and after.
- **Wording only:** 「the card」 / 「this card」 stands on many comment
lines in this package. It carries no number, and neither instrument sees
it. This diff removes one antecedent of it, and the rewrite repairs that
one in place (`protocol.ui-view-hidden-columns.test.ts:46`).

---
_Generated by [Claude
Code](https://claude.ai/code/session_017xfMoEjKUuSh2xYB8sCozp)_

Co-authored-by: Claude <noreply@anthropic.com>

Author: objectstack-fleet[bot]

packages/metadata-protocol/src/discovery-schema-conformance.test.ts

@@ -303,7 +303,7 @@ describe('[#4828] getDiscovery() conforms to DiscoverySchema', () => {
       expect(DiscoverySchema.safeParse(discovery).success).toBe(true);
     });
 
-    // [#6287] `preview` left this list when it stopped being unrecognised: it is
+    // [commit 84c86fb45] `preview` left this list when it stopped being unrecognised: it is
     // a declared `EnvironmentTypeSchema` member and now has a stated fold
     // (`sandbox`), so asserting `development` for it here would assert the
     // opposite of the mapper's decision. The RULE these rows pin — an unknown

packages/metadata-protocol/src/get-meta-item-org-read-gate.test.ts

@@ -7,7 +7,7 @@
  *
  * ── The defect, and why the singular verb is the sharper half ─────────────
  *
- * Commit 96326040f (PR #14767) moved {@link organizationIdForMetaRead} INSIDE the PLURAL
+ * Commit 96326040f moved {@link organizationIdForMetaRead} INSIDE the PLURAL
  * verb, `getMetaItems`, and deliberately did not carry to this one. There, the
  * two `queryByOrg` reads are UNIONed, so an ungated organization can only ADD
  * rows — the resurrection that card is about. Here the two `findOverlay` reads

packages/metadata-protocol/src/get-meta-items-org-read-gate.test.ts

@@ -121,7 +121,7 @@ function makeHarness(rows: StoredRow[]) {
                     return (r as unknown as Record<string, unknown>)[k] === v;
                 }),
             );
-            // `check:objectql-double-limit` (#10978) — hold the caller's bound,
+            // `check:objectql-double-limit` (commit 4c9780c7a) — hold the caller's bound,
             // applied AFTER the filter and BY PRESENCE.
             return opts?.limit === undefined ? matched : matched.slice(0, opts.limit);
         },

packages/metadata-protocol/src/meta-overlay-cache.test.ts

@@ -176,7 +176,7 @@ function makeHarness(rows: StoredRow[], options: HarnessOptions = {}) {
                     return (r as unknown as Record<string, unknown>)[k] === v;
                 }),
             );
-            // `check:objectql-double-limit` (#10978) — hold the caller's bound,
+            // `check:objectql-double-limit` (commit 4c9780c7a) — hold the caller's bound,
             // applied AFTER the filter and BY PRESENCE. A double that hands
             // back every row it matched cannot tell a dropped bound from no
             // bound at all.

packages/metadata-protocol/src/migrations/read-probe.test.ts

@@ -30,7 +30,7 @@
  * instead. The MySQL arm runs against a real server in
  * \`seed-tenancy-backfill.live-mysql.test.ts\`; the SQLite arm end to end against
  * a real \`SqlDriver\` in \`packages/runtime\`'s
- * \`seed-tenancy-autonumber-split.integration.test.ts\`; and [#17621] the POSTGRES
+ * \`seed-tenancy-autonumber-split.integration.test.ts\`; and [commit 7e74af3df] the POSTGRES
  * arm — recorded here as NOT MEASURED anywhere for as long as that was true —
  * now runs in \`seed-tenancy-backfill.live-postgres.test.ts\`, on the postgres
  * service the \`Temporal Conformance\` job already provisions. All three arms are

packages/metadata-protocol/src/migrations/seed-tenancy-backfill.live-postgres.test.ts

@@ -1,7 +1,7 @@
 // Copyright (c) 2026 ObjectStack. Licensed under the Apache-2.0 license.
 
 /**
- * #17621 — the seed-tenancy backfill's statements, RUN on a live PostgreSQL.
+ * Commit 7e74af3df — the seed-tenancy backfill's statements, RUN on a live PostgreSQL.
  *
  * ## The gap this closes
  *
@@ -52,7 +52,7 @@
  * uses) and reports a named SKIP without one — never a silent pass. A runner
  * that knows it provisioned the server sets `OS_EXPECT_LIVE_DIALECT_MATRIX=1`,
  * which turns the missing URL into a failure, so a dropped `env:` line cannot
- * quietly return this seam to the no-coverage state #17621 records.
+ * quietly return this seam to the no-coverage state commit 7e74af3df records.
  *
  * ## Isolation — a per-file SCHEMA, and why the resolver is the MySQL-named one
  *
@@ -253,7 +253,7 @@ describe.skipIf(!PG_URL)('#17621 seed-tenancy backfill on a LIVE PostgreSQL', ()
     const statements: Array<[string, string, unknown[]]> = [
       ['presence probe (fallback arm)', buildSequencesPresenceSql(c), []],
       // The statement the boot path actually runs now. It is in this list for
-      // the reason every other one is here, and #17621 exists because it was
+      // the reason every other one is here, and commit 7e74af3df exists because it was
       // the ONE arm this list could not previously contain.
       ['presence probe (catalog arm)', buildTablePresenceSql(SEQUENCES_TABLE, c) as string, []],
       ['split probe', buildSplitProbeSql(c), [GLOBAL_TENANT, GLOBAL_TENANT]],

packages/metadata-protocol/src/protocol-publish-drafts-closure.test.ts

@@ -35,7 +35,7 @@
  * The `objects` and `permissions` gaps are the same defect at advisory
  * severity: they do not refuse, they manufacture findings that describe
  * nothing. Both are pinned below, because a closure that is uniform is the
- * property #9612/#10058 declared and a per-collection patch is what produced
+ * property #9612/commit 6f5a44976 declared and a per-collection patch is what produced
  * this card.
  *
  * ## The discriminating tests are the ones that expect SILENCE

packages/metadata-protocol/src/protocol-publish-drafts-object-field-refs.test.ts

@@ -244,7 +244,7 @@ describe('publishPackageDrafts refuses a dangling object field-name list (#15254
         expect(causal.code).toBe('INVALID_METADATA');
 
         // …with a rule id and the offending path, which is what has to reach
-        // the author's screen. The path is name-keyed on the wire (#10064) —
+        // the author's screen. The path is name-keyed on the wire (commit def0d3e63) —
         // `objects.<name>.<key>[i]`, never the gate's private snapshot index.
         const wire = JSON.stringify(causal);
         expect(wire).toContain('object-field-ref-unknown');

packages/metadata-protocol/src/protocol.batch-row-http-status.test.ts

@@ -543,7 +543,7 @@ describe('[#14723] section 7 — a `DuplicateRecordError` row reports `UNIQUE_VI
     });
 
     it('[GUARD] a producer that merely SPEAKS `DUPLICATE_RECORD` is not the engine\'s envelope and keeps its own code', async () => {
-        // The same discrimination the whole-request arm makes (#14389 §5):
+        // The same discrimination the whole-request arm makes (commit 10220a7bf):
         // the gate is the registered code AND the class name. A hook throwing
         // the registered member from its own body is a different producer
         // speaking a member of the vocabulary; the verbatim rule still applies.

packages/metadata-protocol/src/protocol.datasource-delete-prolongation.test.ts

@@ -17,7 +17,7 @@
  * re-verification carrier: re-measure the prolongation once that fix lands, and
  * close only on the measurement.
  *
- * The re-verification (PR #14431) found that #13331 closes the registry-heal
+ * The re-verification (commit a98b61b3e) found that #13331 closes the registry-heal
  * half but opens a SECOND gap on the read-in-window path (see "Arms" below,
  * preserved as history). Ruling A′ (2026-09-03, director batch #22, comment
  * 5528370129) ordered the fix for that gap: `applyRemoteMetadataMutation`
@@ -110,7 +110,7 @@
  *      writes the deleted row straight back into the registry the bridge
  *      just healed — and the registry has no TTL.
  *
- * So the prolongation was — PR #14431's finding, kept as history below —
+ * So the prolongation was — commit a98b61b3e's finding, kept as history below —
  * neither unconditionally bounded nor unconditionally unbounded, and the
  * discriminator was not time but TRAFFIC:
  *
@@ -132,7 +132,7 @@
  * one. The bounded arm is the quiet-replica arm.
  *
  * ---------------------------------------------------------------------------
- * Why the re-hydration USED TO happen, precisely (history — PR #14431)
+ * Why the re-hydration USED TO happen, precisely (history — commit a98b61b3e)
  * ---------------------------------------------------------------------------
  * Nothing on the bridge's receipt path touched the WRITE EPOCH that keys the
  * overlay cache. `applyRemoteMetadataMutation` re-read the row and repaired the
@@ -180,7 +180,7 @@
  *     re-derived (the dispatch forbids re-deriving that finding).
  *  3. list-cache TTL (#5109) ............. RE-MEASURED, then CLOSED. At this
  *     door the cache is `meta-overlay-cache`, not `MetadataManager.listCache` —
- *     PR #14431 found it did not merely delay the correct answer, it FED the
+ *     commit a98b61b3e found it did not merely delay the correct answer, it FED the
  *     untimed registry, converting a 30s residue into an unbounded one on any
  *     read landing in the window. Ruling A′'s bump on the receipt path retires
  *     this cache at the moment of convergence, so the read that follows has
@@ -691,7 +691,7 @@ describe('[#13609] ⭐ the re-verification: how long does the peer keep serving
 
         // `measureProlongationMs` reads the door once BEFORE advancing the
         // clock — i.e. inside the residue window, which is what a replica
-        // under load does continuously. Previously (PR #14431) that read
+        // under load does continuously. Previously (commit a98b61b3e) that read
         // converted the bounded cache residue into an unbounded registry
         // entry. Now the convergence bump has already retired the cache
         // before this first read runs, so the read-in-window traffic