Cross-seat transfer from the domain:metadata/engine-core seat (PM session session_01XeQRiAa7vYRVX5Fog7Zby8). Filed here rather than kept in-lane because it lands in packages/spec + @objectstack/lint, and packages/spec routes to the domain:spec seat regardless of who needs it.
Source: #8690, and specifically its maintainer ruling 5299879288 (delegated adjudication; delegation 2026-08-15 verbatim 「决策你直接帮我做」, batch confirmed 「同意」). This is not a new proposal — the decision is already made and this card exists only to put the ruled work in the right seat's queue at the right tier.
What was ruled
Ruled: Option B as the contract answer, with Option C shipped alongside. Explicitly not A.
C — refuse the declared preset vocabulary (last_7_days/last_30_days/last_90_days as bare strings in a temporal filter) at publish time, packages/spec + @objectstack/lint. Strongest AI-error-resistance win, ships even though not sufficient alone.
This card is the C half only. The B half (refuse at the ObjectQL engine's filter collection point, plus a NativeSQLStrategy.canHandle decline via a new StrategyContext hook) is dispatched separately on #8690 and lands in packages/objectql + service-analytics. They are two refusals at two boundaries, not one refusal split across packages — the ruling lists them as separate deliverables, and the ruling's own ⛔ against "splitting the refusal across packages" applies within each half, not between them.
⚠️ Mandatory model tier
The ruling's closing line: "C narrows an accept set on a published authoring surface ⇒ claude-fable-5 tier mandatory for the C half." That is the standing non-discretionary clause, not a suggestion. The only sanctioned downgrade is the maintainer's 2026-08-13 quota exemption (「fable 如果用完了,可以用 opus」) — and only when fable is measurably unavailable, with the tier and reason recorded in the claim comment.
The defect this closes
last_7_days / last_30_days / last_90_days are real, declared preset names in the dashboard schema. The shipped console lowers them to {N_days_ago} / {today} macros before they reach the query API, so the console path is safe. Any other caller that has the preset name in hand — a saved report, an integration, an MCP client, an AI-authored query — sends the name itself.
Measured end-to-end on InMemoryDriver with a declared datetime field, 51 rows seeded / 38 in-window:
$gte "last_30_days" HTTP 200 count=0 <- the defect: silent zero
$gte "not-a-date-at-all" HTTP 200 count=0
$gte "{30_days_ago}" HTTP 200 count=38 <- positive control
$gte "{TODAY}" REFUSED code=FILTER_TOKEN_UNKNOWN status=400
$gte "{not_a_token}" REFUSED code=FILTER_TOKEN_UNKNOWN status=400
So the vocabulary is declared in one layer and unrecognised in the next, with no error at the boundary — a {placeholder} the resolver does not know is rejected loudly with a list of resolvable tokens, while a declared preset name is bound as-is and compares false against every row.
C is the half that closes this at authoring time, where an AI-generated dashboard is actually produced, rather than tolerating it at consume time. It does not close the ad-hoc integration/MCP path — that is what B is for.
⛔ Scope boundary carried from the ruling
The empty-string cell stays its own card — B and C scope to non-empty strings and must not decide it in passing.
Measured: $gte "" binds as '', and because every canonical UTC text sorts at or above it, the predicate is satisfied by every non-null row — 51 of 51. (#8690's body table records 38 for that row; that is a transcription error. Its prose, "silently drops the constraint and returns everything", is the correct account.) Leave that cell exactly as it is.
Why this is repair, not a feature
By the mechanical boundary test the accept set moves inward, restoring declared = enforced: a vocabulary the platform itself declares becomes refused at the layer that cannot interpret it, instead of silently producing an empty result. The loud path already exists one branch over (FILTER_TOKEN_UNKNOWN), so this reuses a settled refusal identity rather than minting one.
Suggested verification
Pin that a bare preset name in a temporal filter is refused at publish with a message naming the macro spelling that does work ({30_days_ago}), and keep a discriminating control in the same test — a legitimate temporal comparand still publishing cleanly. A refusal pin with no positive control cannot show the rule is selective rather than blanket.
Origin: carved out of #8690 by the dispatching seat under that card's ruling, so the mandatory tier and the packages/spec ownership rule are both honoured rather than folded into an opus engine-core dispatch.
Cross-seat transfer from the
domain:metadata/engine-coreseat (PM sessionsession_01XeQRiAa7vYRVX5Fog7Zby8). Filed here rather than kept in-lane because it lands inpackages/spec+@objectstack/lint, andpackages/specroutes to thedomain:specseat regardless of who needs it.Source: #8690, and specifically its maintainer ruling
5299879288(delegated adjudication; delegation 2026-08-15 verbatim 「决策你直接帮我做」, batch confirmed 「同意」). This is not a new proposal — the decision is already made and this card exists only to put the ruled work in the right seat's queue at the right tier.What was ruled
This card is the C half only. The B half (refuse at the ObjectQL engine's filter collection point, plus a
NativeSQLStrategy.canHandledecline via a newStrategyContexthook) is dispatched separately on #8690 and lands inpackages/objectql+service-analytics. They are two refusals at two boundaries, not one refusal split across packages — the ruling lists them as separate deliverables, and the ruling's own ⛔ against "splitting the refusal across packages" applies within each half, not between them.The ruling's closing line: "C narrows an accept set on a published authoring surface ⇒
claude-fable-5tier mandatory for the C half." That is the standing non-discretionary clause, not a suggestion. The only sanctioned downgrade is the maintainer's 2026-08-13 quota exemption (「fable 如果用完了,可以用 opus」) — and only when fable is measurably unavailable, with the tier and reason recorded in the claim comment.The defect this closes
last_7_days/last_30_days/last_90_daysare real, declared preset names in the dashboard schema. The shipped console lowers them to{N_days_ago}/{today}macros before they reach the query API, so the console path is safe. Any other caller that has the preset name in hand — a saved report, an integration, an MCP client, an AI-authored query — sends the name itself.Measured end-to-end on
InMemoryDriverwith a declareddatetimefield, 51 rows seeded / 38 in-window:So the vocabulary is declared in one layer and unrecognised in the next, with no error at the boundary — a
{placeholder}the resolver does not know is rejected loudly with a list of resolvable tokens, while a declared preset name is bound as-is and compares false against every row.C is the half that closes this at authoring time, where an AI-generated dashboard is actually produced, rather than tolerating it at consume time. It does not close the ad-hoc integration/MCP path — that is what B is for.
⛔ Scope boundary carried from the ruling
Measured:
$gte ""binds as'', and because every canonical UTC text sorts at or above it, the predicate is satisfied by every non-null row — 51 of 51. (#8690's body table records 38 for that row; that is a transcription error. Its prose, "silently drops the constraint and returns everything", is the correct account.) Leave that cell exactly as it is.Why this is repair, not a feature
By the mechanical boundary test the accept set moves inward, restoring
declared = enforced: a vocabulary the platform itself declares becomes refused at the layer that cannot interpret it, instead of silently producing an empty result. The loud path already exists one branch over (FILTER_TOKEN_UNKNOWN), so this reuses a settled refusal identity rather than minting one.Suggested verification
Pin that a bare preset name in a temporal filter is refused at publish with a message naming the macro spelling that does work (
{30_days_ago}), and keep a discriminating control in the same test — a legitimate temporal comparand still publishing cleanly. A refusal pin with no positive control cannot show the rule is selective rather than blanket.Origin: carved out of #8690 by the dispatching seat under that card's ruling, so the mandatory tier and the
packages/specownership rule are both honoured rather than folded into an opus engine-core dispatch.