Pin cross-site unprovisioned-anchor content invariant (wording convergence blocked by drift)

[qq9340100]

Part of #8405 — #8405 is NOT fixed by this PR; it remains open pending a maintainer decision (see below).

What this is

#8405 asked to converge warnUnprovisionedAnchors (validate-expressions.ts) and
unprovisionedPointer (validate-semantic-roles.ts) onto the shared
unprovisionedAnchorCause/unprovisionedAnchorHint builders in system-fields.ts
(added by #8340), with byte-identical output as the acceptance bar — this is a
de-dup, not a rewording.

Diffing the emitted text programmatically (not by eye) against real sample args shows
both sites already drifted from the shared builders, and from each other, before
this PR touched anything:

• validate-expressions.ts's cause clause: 'FIELD', an injected system column…
  (comma) vs. the shared unprovisionedAnchorCause()'s 'FIELD' is an injected system column… (space + "is"). The rest of the cause sentence is byte-identical.
• validate-expressions.ts's hint clause: carries this column, declare 'FIELD' in the object's own fields… vs. the shared unprovisionedAnchorHint()'s carries 'FIELD', declare it in OBJECT_NAME's own fields… — the shared builder
  interpolates the real object name; the site hard-codes the generic "the object's".
• validate-semantic-roles.ts's cause clause is a materially different sentence:
  double quotes instead of single, "no storage" instead of "NO storage", a
  differently-ordered reason clause ("…but the remote schema owns the table and no column backs it. Every consumer renders it empty on every record."), none of which
  matches the shared builder's text.
• validate-semantic-roles.ts's hint clause ends "…otherwise point ${slot} at a real remote column" instead of the shared builder's ownership/systemFields opt-out
  language.

So the issue's own premise — "Not a defect today — the three copies currently
agree" — does not hold at the byte level this card's ruling requires. Wiring
either site to call the shared builders as written would change its emitted
diagnostic text for existing consumers, which is exactly what the ruling says to stop
and report instead of shipping under a de-dup card. See the #8405 dev report
(issue comment) for the full character-level diff evidence.

What this PR actually does

No production code changes. It adds the one thing that survives regardless of how
the wording question gets resolved: a cross-site content-invariant test
(validate-expressions.test.ts, describe('cross-site unprovisioned-anchor convergence (#8405)', …)) pinning that both sites report the same object, same
column, and the same ADR-0015 reason even though their exact phrasing differs today.
A future edit to either site can drift in wording freely but can no longer silently
drop the object name, the field name, or the ADR-0015 citation that makes the finding
actionable.

Decision needed (see dev report open_questions)

Real convergence of the exact wording needs a deliberate decision, not a silent pick
under this card:

• Recommended: a dedicated wording-alignment PR (with a changeset, since it
  changes shipped lint diagnostic text) that adopts the shared system-fields.ts
  builders as canonical everywhere and updates both originals' literal text to match.
• Alternative: widen the shared builders to parametrize the differences so each site
  keeps its current exact text — not recommended, since it institutionalizes what
  looks like accidental drift as permanent, three-way design.
• Alternative: do nothing further and rely only on the content-invariant pin added
  here — cheapest, but leaves the "three wordings of one verdict" trust gap finding: #8116's two unprovisioned-anchor diagnostics still hand-copy the cause sentence #8340 centralized into system-fields.ts #8405
  itself called worse than cosmetic.

skip-changeset: no production code changed, test-only addition, releases nothing.

Generated by Claude Code

---

Generated by Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
objectstack	Ignored		Aug 13, 2026 4:46pm

[github-actions]

📓 Docs Drift Check

No hand-written docs reference the 0 changed package(s). ✅