fix(spec,service-analytics): the matrix's token axis + its sixth consumer, and the dropped $between it found (ADR-0053 D-A3.1) - #4109
Merged
Conversation
…atetime backends (ADR-0053 D-A3) (#4081) The temporal seam broke four times (#3650 window dropped entirely, #3773 epoch-as-Julian-day NULL buckets, #3777 bare-day upper bound losing the final day, #4047 type-bracket comparison returning nothing on memory/mongo), and each break was found by accident: every fix left a suite proving its own issue against its own fixture, with no shared standard to hold the six evaluation surfaces together. This adds the standard — the temporal twin of filter-logic-conformance (#3774), and the one ADR-0053 decision (D-A3) that had never been built: - `packages/spec/src/data/temporal-conformance.ts` exports `TEMPORAL_CONFORMANCE_ROWS` (one fixture spanning exact-midnight, intra-day, next-midnight, month-end last-millisecond, leap-day and pre-epoch instants, each tagged with a writer form for the D-E4 mixed-writer axis) and `TEMPORAL_CONFORMANCE_CASES` ({ name, fieldType, operator, filter, tokenFilter?, dateRange?, expected, note }) — field-type × operator × bound-semantics (D-D2) × relative-token cells, asserting row-id sets, never emitted SQL. Each note names the incident the case guards. - Six thin consumers: driver-sql (canonical + un-backfilled legacy storage; live PG/MySQL via the existing CI temporal job), driver-sqlite-wasm (inheritance guard), driver-memory and driver-mongodb (mixed-writer seeds), formula's matchesFilterCondition (RLS write-side check), and the analytics preview evaluator (where + timeDimensions.dateRange, the #3650 surface). - The token axis runs end-to-end: consumers that can reach @objectstack/core resolve each case's tokenFilter/dateRange against the pinned TEMPORAL_CONFORMANCE_NOW and must land on the same rows as the literal spelling, so resolver drift and evaluator drift are distinguishable at a glance. Deliberate scope, same discipline as the filter-logic table: a case belongs here only if every backend must agree on it. The schema-aware-only cells (bare-day $eq/$in on datetime, $gt at an exact stored midnight, Field.time) stay pinned in the per-driver suites and are documented as exclusions. Closes #4081. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TqqZmPS5a4gJGBoCTwipFr
|
The latest updates on your projects. Learn more about Vercel for GitHub. 1 Skipped Deployment
|
…n top Main landed its own D-A3 matrix (#4098: TEMPORAL_ROWS/TEMPORAL_CASES + the preview $between fix) while this branch was in flight. Resolution keeps main's landed table, naming and consumers as the base, and re-applies this branch's increments — exactly the coverage #4098's ADR addendum recorded as still open: - the relative-token axis: TemporalCase.tokenFilter/dateRange + TEMPORAL_NOW, resolved via @objectstack/core resolveFilterTokens in the five core-reachable consumers (formula stays literal-only by dependency stance) - TemporalRow.writerForm: mixed writer-population seeding (D-E4) in the sql/wasm/memory/mongo consumers - a pre-epoch fixture row, and the date equality/$in cells (#1874's shape) - a driver-sqlite-wasm consumer pinning the inherited SqlDriver seam - a legacy-storage sweep in the driver-sql consumer (un-backfilled mixed epoch/naive column through the read-repair path) The duplicate preview consumer from this branch is dropped in favour of main's preview-temporal-conformance.test.ts, extended with the token and timeDimensions.dateRange sweeps; this branch's changeset is rewritten as the increment (.changeset/temporal-conformance-token-axis.md) alongside #4098's. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TqqZmPS5a4gJGBoCTwipFr
github-actions
Bot
added
documentation
Contributor
📓 Docs Drift CheckThis PR changes 2 package(s): 108 hand-written doc(s) reference the affected code and may need an implementation-accuracy re-verification:
|
os-zhuang
changed the title
…$between it found (ADR-0053 D-A3.1) NativeSQLStrategy is the surface #3650 was actually about, and it was listed in the conformance matrix's own backend table with no consumer — because every existing suite for it asserts the emitted SQL string, and a DROPPED predicate is invisible to a string assertion: the SQL stays valid, just wider. Executing the shared cases against a real engine (sql.js, the same pure-WASM engine driver-sql falls back to) and asserting row ids is what D-A3 asked for, and it immediately showed `$between` returning the entire table. The cause sat one layer below the strategy, in the shared filter-normalizer: `$between` was absent from MONGO_TO_CUBE_OP, and an unmapped operator is `continue`d, so the predicate vanished from the compiled WHERE clause. Both strategies read that normalizer, so the ObjectQL aggregate path was affected too. User-visible symptom: a widget with a range filter charts the whole dataset, with nothing in the SQL to suggest a filter was ever requested. `$between [min, max]` now LOWERS to its two bounds (gte + lte) rather than gaining an operator of its own, so a range's max inherits the calendar-day whole-day rule (#3777) from each strategy's existing upper-bound handling — NativeSQLStrategy compiles a bare-day upper bound half-open itself, the ObjectQL path gets the same rule from the driver — instead of needing a second implementation to keep in step. That is how #4098 closed the same defect on the preview evaluator. A malformed `$between` throws rather than being dropped, the stance driver-memory took for the same shape in #3948. The consumer runs the literal, relative-token and timeDimensions.dateRange spellings of every shared case. Rows are seeded canonical, where the driver's temporal hooks are identities, so the context omits them and exercises the same "absent = identity" path Postgres takes (D-A2); the un-backfilled mixed column stays covered where the driver truth lives. The cause itself is NOT fully closed: `$startsWith`, `$endsWith`, `$null` and `$regex` are still silently dropped by the same `continue`. Filed as #4128 rather than expanded into this change, with the #3948 precedent for turning the fallback into a throw. service-analytics 391 green (was 360; +31 from the new consumer). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TqqZmPS5a4gJGBoCTwipFr
os-zhuang
changed the title
The module doc described the shape it produces but never what it actually maps — and in a normalizer, that omission is load-bearing: an unmapped operator is `continue`d, which does not "not support" the predicate, it DELETES it, widening the query to rows the author excluded. That is how the $between defect survived (see #4128 for the four operators still in that state, and the #3948 precedent for making the fallback throw). Mirrors the scope comment the preview evaluator carries for the same reason. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TqqZmPS5a4gJGBoCTwipFr
…query (#4128) Closes the CAUSE behind the $between defect, not just that instance. `normalizeAnalyticsFilters` skipped any operator missing from its map, and a skipped predicate does not narrow a query — it widens it: the SQL stays valid and returns rows the author excluded. Four operators from the spec's authorable vocabulary sat in that state, and a fifth was mapped wrongly: - $startsWith / $endsWith were dropped. Both strategies now compile them — anchored LIKE on the raw-SQL path, the canonical operators (which every driver implements directly) on the ObjectQL path, so an anchored match does not depend on regex dialect. - $null was dropped. It is what the console emits for an "is empty" filter, so such a widget showed every row. - $exists was mapped value-INDEPENDENTLY to `set`, so {$exists: false} compiled to IS NOT NULL — the inverse of what it asks. It and $null are now resolved explicitly: a key→name map cannot express an operator whose meaning flips with its value, which is exactly how that inversion got in. - $notContains reached ObjectQLStrategy, which had no arm for it and fell to a `default` returning a bare value — compiling "does not contain x" as "equals x". - Unknown operators now THROW on both surfaces rather than being dropped (normalizer) or reinterpreted as an equality (ObjectQL). #3948's call for the same shape. $or / $not remain skipped — expressing them needs a recursive WHERE builder rather than the flat array the strategies consume. That gap is declared in the module doc rather than silent. filter-operator-coverage.test.ts runs the whole vocabulary against a real SQLite engine and asserts ROW IDS; six of its cases fail without this change. A dropped predicate is invisible to the SQL-string assertions the strategies' other suites use, which is how these survived. service-analytics 413 green. Closes #4128. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TqqZmPS5a4gJGBoCTwipFr
This was referenced Jul 30, 2026
Contributor
Author
|
CI 说明：Lint & Type Check 的红不是本 PR 引入的，在 base 分支上同样失败。 失败的其实不是 ESLint 步骤，而是同一 job 里的 在纯净的 成因是两个 PR 的竞态：#4112 删掉了那个挂载，#4116 在守卫账本里声明了它，先后合入未互相 rebase。已单独立项 #4140（含复现步骤与建议的销账式修法）。 本 PR 的其余检查照常，待 main 恢复后我会重跑并推进合并。 Generated by Claude Code |
…rature-consistency-akyja5
os-zhuang
marked this pull request as ready for review
July 30, 2026 11:26
os-zhuang
pushed a commit
that referenced
this pull request
Jul 30, 2026
The temporal conformance matrix (#4109) gives every case a `tokenFilter` spelling, and the driver + analytics suites resolve it before asserting. This suite runs only the literal `filter`, with nothing saying why — so the absence reads as an omission, and the next person either "fixes" a non-gap or assumes coverage that cannot exist. It is architecture: an RLS `check` is a CEL expression (`check` is a string, compiled by rlsCompiler.compileFilter) where a relative date is the function today() evaluated at compile time. A {token} string never reaches this evaluator — resolveFilterTokens runs on the ObjectQL read path and the write-side check does not go through it; verified that plugin-security never calls the resolver. Comment only; the suite is unchanged and still 13/13. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01EkL3RGJrzjLEGURsLxfS2f
This was referenced Jul 30, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
#4081（ADR-0053 D-A3）的补完。矩阵本体已由 #4098 先行落地；本 PR 与其合流，补上 #4098 的 ADR 附录里明确记为仍然开放的两处覆盖，并修掉补覆盖时矩阵当场抓到的缺陷——以及顺着成因挖出的一整类。
一、relative-token 轴（#4098 附录记为开放）
四次事故都发生在 token 解析之下，但作者实际书写的就是
{today}/{90_days_ago}/{current_month_end}，而没有任何测试证明解析后的比较值在各后端落到同一批行。TemporalCase新增tokenFilter?与dateRange?，加上钉死的参考瞬间TEMPORAL_NOW。@objectstack/core的消费者把 token 拼写解析后必须落在与字面量拼写相同的行集——resolver 漂移与 evaluator 漂移由此可分。formula依赖刻意止于 spec，只跑字面量。TemporalRow.writerForm（D-E4 混合写入者播种）、负 epoch 行、date的相等与$in（[P2] Provide a declarative time-relative trigger (avoid fragile date-equality on record-change) #1874 起源形态）、driver-sqlite-wasm继承护栏、driver-sql未回填 legacy 双扫。二、第六个消费者：
NativeSQLStrategy它正是 #3650 事发的那个面，也是唯一手工拼 SQL、不走驱动编译器的后端，却在矩阵的后端表里没有消费者。原因很说明问题：它现有的套件断言的都是 SQL 字符串，而谓词被丢掉对字符串断言完全隐形——SQL 依然合法，只是更宽。新消费者用真实引擎（
sql.js）执行并断言行 id。三、它当场抓到的缺陷：
$between谓词凭空消失首跑即红：
$between返回全表。成因在共享的filter-normalizer——MONGO_TO_CUBE_OP里没有$between，未映射的算子被continue掉，谓词直接从 WHERE 消失。两条策略都吃这个归一化器，所以都中招。用户侧症状就是 #3650：带区间过滤的图表画出整个数据集。修法：
$between下降为它的两个界（gte+lte），上界自动继承各策略已有的整日日历规则，不需要第二份实现——与 #4098 修 preview 的做法一致。四、收口成因本身（Closes #4128）
只修
$between等于只修实例。按同一把尺子把整个可授权词汇过了一遍，又挖出两处比「丢弃」更糟的——语义被悄悄写反：$startsWith/$endsWithLIKE 'x%'/LIKE '%x'；ObjectQL 侧透传规范算子，锚点不依赖 regex 方言$nullIS NULL/IS NOT NULL$existsset，{$exists: false}编译成IS NOT NULL——与语义完全相反$null一起显式解析：键→名的映射表表达不了「含义随取值翻转」的算子，这正是反转的成因$notContainsObjectQLStrategy没有 case，落到default: return v0返回裸值——「不包含 x」被编译成「等于 x」$notContainscontinue丢弃；ObjectQL 层当成相等$or/$not仍跳过（需要递归 WHERE 构建器），但这一条现在写进模块文档作为已声明的缺口，不再是隐性的。验证（本地全绿）
两个新套件都用
git stash实测过修复前的行为：$between4 例红、算子覆盖 6 例红。mongodb 套件本沙箱无二进制而 skip，CI 真实执行。spec 八个生成物 gate 全 PASS。Refs #4081，Closes #4128。
🤖 Generated with Claude Code
https://claude.ai/code/session_01TqqZmPS5a4gJGBoCTwipFr