Status: fixed-model generated cache execution and ownership checkpoint verified. Full G08/P5 remains partial; this checkpoint does not resolve the separate quoted-Line attachment question or qualify all cache/concurrency/trace configurations.
Runtime v5.0.6 fe82a18982415f1746df9a2cf7c42f5e9abdc6d3 and producer
45e1572c4956ca963410d8ea436ead3c8857087a remain locked. The unchanged complete
generated dynamic_json_rust crate supplies actual Q methods and execution.
Recording backend wrappers delegate get/put/invalidate calls to the original and
C++ InMemoryAggregationCache implementations; they do not replace caching logic.
All database operations use the existing fixed FK-enabled SQLite fixture.
The source fetch_prepared_all_with_aggregates takes query-owned relation aggregate declarations out of the base execution query, then merges explicit declarations without adding identical duplicates. Source query-owned duplicates remain; only duplicate explicit transport is excluded. C++ EntityRowsOperation now performs this normalization at its owned snapshot boundary. It preserves the reusable caller query and existing independent service/executor/provider roles.
This fixes a real semantic difference: C++ previously retained child aggregation metadata in the base-row cache key. Changing a child filter then missed the base cache and issued extra SQL, unlike the fixed runtime. The fix does not rewrite cache observations or omit key fields in the canonicalization function.
The source matrix observes literal cache keys, namespace, requested max age, hit/miss, put and invalidation events, actual typed relation metric values, and cumulative SQL/parameters/comment/purpose. It covers count and sum, absent cache backends, root/child enablement, TTL changes, default and explicit propagation, explicit child overrides, empty roots, comments, purpose, filters and clear. Direct SQLite updates demonstrate stale cached values until expiry or clear; governed generated saves exercise namespace invalidation independently.
Zero TTL means no expiry. Four real-clock scenarios use a wide 30 ms separation against a 1 ms maximum age; they do not claim exact scheduling at the expiry boundary. Cache hits return copied rows and never synthesize SQL-success logs. The source's purpose-sensitive keys and other actual key rules are retained.
Context tests check independent get results, namespace invalidation, explicit Global cache ownership after source UserContext destruction, destination-scope nested payloads, retained results after invalidation, and repeated actual SQLite execution with identical core/explicit aggregate transport. A null default PMR resource ensures allocator propagation. Global caches are application-owned resources, never hidden singletons or implicit ownership extensions.
Artifacts:
- tests/parity/rust_field_layout/src/bin/generated_aggregation_cache_5_0_6.rs
- tests/parity/generated_aggregation_cache_probe.cpp
- tests/parity/generate_generated_aggregation_cache.py
- tests/parity/compare_generated_aggregation_cache.py
- tests/parity/generated_aggregation_cache_5_0_6_cases.json.gz
- teaql_runtime/tests/generated_aggregation_cache_tests.cpp
The source corpus contains 204 exact scenarios and 588 actual generated SQLite executions, including 20 governed-save scenarios, literal event/key comparisons, missing-backend behavior and four expiry/no-expiry cases. The mutation uses a typed Decimal value in both APIs, preserving SQL parameter provenance without rewriting logs. All observations match without normalization. The observed events include 756 get calls: 394 hits and 362 misses, with 362 puts and 20 entity-namespace invalidations. Native CTest passes 5/5, targeted ASan/UBSan 3/3 with leak scanning disabled, and actual module facet/relation execution 1/1. Formatting checks pass for 928 handwritten C++ files; locked reference/export/path integrity and diff checks pass. All 17 capability records remain unqualified. No generator or application API changed, so Java/compiler/reproduction gates from the prior checkpoint were not rerun. The pinned SQLite archive is not sanitizer instrumented. No MIPS, physical-router or GPU adapter execution is claimed. Full runtime-scope telemetry, privacy/trace provenance, replacement while pending, additional model/nested configurations and complete cache governance remain separate qualification work.