Third `type === 'objects'` tolerance in protocol.ts: `applyObjectRegistryMutation` CONSUMES the tolerant spelling rather than only guarding on it

[hotlong]

Blocked-by: #8819

Observation recorded while implementing #8820. ⚠️ Reachability is NOT measured — filing it plainly rather than grading it myself, because #8820 is precisely the card where an unmeasured reachability claim was the whole defect.

What was measured

At origin/main b69d0f5e4 there are three type === 'objects' tolerances in packages/metadata-protocol/src/protocol.ts:

line	site	what it does with the tolerant value
:11487	ensureObjectStorage	guard only — type never used again
:11542	dropObjectStorage	guard only — type never used again
:10869	applyObjectRegistryMutation	consumes it

The first is handled by #8820. The third is different in kind:

private applyObjectRegistryMutation(request: { type: string; name: string; … }): void {
 if (request.type !== 'object' && request.type !== 'objects') return;
 this.engine.registry.registerItem(request.type, request.item, 'name');

It does not merely admit a plural, it registers under it. That is the shape canonicalMetaType's header describes as already having cost the repo a real bug, in its own words:

One plural-spelled read minted a plural registry entry, listItems('actions') stopped being empty, and the singular fallback that had been supplying the code-authored items never ran again — so one overlay row shadowed an entire code-authored listing, and survived the DELETE that was supposed to lift it.

So if any caller can deliver 'objects', this mints exactly that entry. If none can, the limb is dormant and the tolerance is still the spelling-tolerant-lookup shape that header rejects.

Why I did not measure it

Sole caller is applyRegistryWriteThrough (:11193), which has four call sites: :12485, :13348, :15503, :15793. :13348 folds (that is #8820's edit). ⚠️ :15503 and :15793 sit inside the region #8819 holds in flight, so tracing them to their producers was not mine to do this round. Deciding this one needs all four traced to where the value is produced — the step whose omission is the entire subject of #8820.

Suggested disposition

Sequence after #8819 lands, then trace all four callers. If any delivers a plural, this is a live registry-shadowing defect rather than dormant code. Note the fix shape #8820 settled on is available here too and is stronger than tightening the guard: fold at the producer so the tolerant spelling cannot arrive, rather than teaching the predicate which spellings to accept.

Backlink: #8820 (same tolerance family, same file; that card's measurement is what surfaced this one). See also the dropObjectStorage twin at :11542, which #8820's PM notes as looking dormant with one call site traced and one not.

[os-project-manager]

Triage: domain:metadata, pm:blocked, type Task. Body edited by triage to add the machine-readable Blocked-by: #8819 line (content otherwise unchanged, restored from the escaped API read) — the card's own disposition says sequence after #8819 lands, and two of the four caller sites sit inside #8819's in-flight region, so the unlock scan should hand it back automatically.

On unlock: re-verify the file surface on the merged ref first (#8819's landing may fold or move these sites), then trace all four applyRegistryWriteThrough callers to their producers before deciding live-vs-dormant. Fold-at-producer is the precedented fix shape (#8820) if any caller delivers a plural.

---

Generated by Claude Code

[os-zhuang]

Claim: PM loop round 1
Session: session_01NTKPDRoynY8i3HmdSFUxFj
Branch: claude/issue-8862-object-registry-mutation-plural
Worktree: objectstack-issue-8862
Domain: domain:metadata
File surface (region-level, per this lane's concurrent-editor rule): packages/metadata-protocol/src/protocol.ts — applyObjectRegistryMutation (:11009) and applyRegistryWriteThrough (:11317) and the four call sites, plus whatever producer the trace lands on, plus a pin file. ⛔ Explicitly not publishPackageDrafts / recordMetadataAudit / promoteDraftForPublish. (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: opus. Type Task, repair lane.

Blocker cleared — verified, not assumed

Blocked-by: #8819 is discharged: #8819 is CLOSED completed, PR #8870 MERGED. The card's own disposition ("sequence after #8819 lands, then trace all four callers") is now live.

Serial constraints cleared — read off the ACTUAL diff, not off symbol addresses

protocol.ts has one unmerged editor: PR #8986 (#8908, draft, +209/−6 in this file). I read its file list and hunk ranges rather than inferring from the function it is named for:

• its protocol.ts hunks are at old-line 3423, 10708, 13835, 13916, 14062, 14209, 14287;
• its patch body mentions recordMetadataAudit, publishPackageDrafts, promoteDraftForPublish, migrateStoredMetadata, getEffectiveLock — and mentions applyObjectRegistryMutation zero times and applyRegistryWriteThrough zero times.

⇒ regions are disjoint; this card is the second concurrent editor under the region exemption (cap 3). ⛔ The exemption suspends queueing, not collision-avoidance: merge origin/main before opening the PR, and merge again once #8986 lands.

⭐ Relevant to this card's subject, not just its scheduling: #8986 deliberately excludes manifest-present plurals (objects) from its new gate — stated in its PR body. So it does not touch, fix, or pre-empt the 'objects' question this card is about. The two do not overlap in either region or verdict.

⚠️ STALE PREMISE — the card's central table is now wrong, and in a way that matters

Re-measured on origin/main @ b537855. The card (written at b69d0f5e4) reports three type === 'objects' tolerances. There is now exactly one === 'objects' occurrence in the whole file:

card says	measured on origin/main today
:11487 ensureObjectStorage — guard only	gone — no 'objects' tolerance remains
:11542 dropObjectStorage — guard only	gone — no 'objects' tolerance remains
:10869 applyObjectRegistryMutation — consumes it	alive, now at :11009

⇒ #8820's landing removed the two guard-only siblings, so the "third of three" framing is spent: this is now the only surviving member of the family, which strengthens the card rather than weakening it. ⛔ The dev must not go looking for the other two and conclude the card is stale.

All four caller line numbers rotted — the card's :12485 / :13348 / :15503 / :15793 are now :12625 / :13506 / :15673 / :16014. Anchor on symbols; re-derive these before trusting them.

PM mechanism reading (⚠️ falsifiable — measure it, do not inherit it)

The consuming shape is intact and, read today, slightly sharper than the card states:

applyRegistryWriteThrough (:11317)
 if (request.type === 'object' || request.type === 'objects') {
 this.applyObjectRegistryMutation(request); // request passed through UNCHANGED
applyObjectRegistryMutation (:11009)
 if (request.type !== 'object' && request.type !== 'objects') return;
 this.engine.registry.registerItem(request.type, request.item, 'name'); // registers under the RAW spelling

So the plural is admitted at the write-through and consumed at the mutation — registerItem('objects', …) is what a plural caller mints. That is the exact shape canonicalMetaType's own header blames for a real prior bug (one plural read minting a plural registry entry that shadowed a code-authored listing).

This is a reading, not a ruling. The card's real question is unchanged and is the deliverable: trace all four callers to where the value is produced and decide live-vs-dormant. premise_still_valid: false with a measured "dormant, here is the proof" is a fully valid delivery — ⛔ do not manufacture a fix to have something to show.

If any caller can deliver a plural: the precedented shape is fold at the producer (#8820's settled answer), ⛔ not teaching the predicate more spellings.

---

Generated by Claude Code