Temporal persistence v2#

The implemented design preserves native Temporal.Instant and Temporal.PlainDate values through the existing SQLite query pipeline. It keeps SQL filtering and index use, reconstructs native values, then uses the existing residual filtering, ordering and pagination stages.

The grammar is:

supported value × carrier × operation × legal history × execution boundary

Its selected rules are concrete:

• Use registered global Temporal constructors for brand validation and native reconstruction. Preserve both kind and value; reject missing constructors and other Temporal kinds. Strings remain strings.
• Encode native values consistently across rows, nested data, metadata and replay. Escape an ordinary record's marker field without moving its other JSON fields.
• Separate lossless canonical text from sortable keys. Instant keys retain nanoseconds; PlainDate keys preserve ISO chronological order across calendars and signed years. Equality additionally retains kind and calendar text.
• Share reference/literal representations between runtime SQL and persisted indexes. Exact equality and correlated IN membership remain safe under NOT. Coalesce retains both representations; existing SQL fallbacks and final cleanup remain. Ordered comparisons require compatible same-kind operands.
• Preserve action validation, durable rollback, wrapper receipts/FIFO and the publication-before-durability boundary. Reuse existing oracle owners.
• Retain the narrower remote-subset wire domain and reject permanent input errors before retry admission. Multiprocess native-value support is not added.

The wrapper oracle exposed an additional receiving boundary: Collection index metadata previously JSON-cloned native literals into strings and gave distinct Temporal literals the same signature. The implementation now preserves native literals while copying the expression structure and includes kind/text in the signature. Its test observes metadata, emitted SQL and actual index use.

The original implementation failed 33 targeted expanded witnesses before the final routing and scalar-index additions. A deliberately wrong lexical-order implementation fails eight cases while still reconstructing native values correctly. The routing witnesses also catch a request that becomes remote during hydration; valid requests now dispatch exactly once. Small and large membership indexes retain SQLite compatibility. The repaired affected suite reports 471 passed, 1 existing TODO, 0 failed. TypeScript and lint pass; lint retains existing require-await warnings. Full results, source hashes and the oracle audit are in evidence.

Old data may be replaced, as the user specified. This additive implementation does not automatically erase it: existing schema-version/reset options can replace already-damaged caches. Empty objects cannot reveal their former type, and older readers cannot interpret the new native encoding.

The claim is bounded to the reached paths and histories. Engine-native constructors, native mobile/browser runs, multiprocess row transport, truncate/delete histories with native values and arbitrary Temporal operations remain unproved. Independently held-out range evidence and formal model checking are absent; host behavior and untested schedules remain a decomposition limit. Index-plan checks establish index use, not a latency guarantee.

The detailed model preserves the initial frozen rules and explicitly dated implementation discoveries. Process records reconstruction, ablation, controls and the projection map. No new lifecycle model or generic codec framework was introduced.