Skip to content

sql_protocol_5_0_6.md

Latest commit

 

History

History
114 lines (96 loc) · 6.93 KB

sql_protocol_5_0_6.md

File metadata and controls

114 lines (96 loc) · 6.93 KB
Raw

SQL request protocol qualification against Rust 5.0.6

This checkpoint advances G09/G13/G14/G17 and P7. Complete capabilities and P0-P7 remain open. The runtime reference is the exact 5.0.6 commit in upstream.lock.json.

Current constructors and preserved historical inputs

The four runnable observers use the actual QueryIntent, MutationIntent, QueryRequest and MutationCommand/MutationRequest contracts. Their independent development workspace is tests/parity/rust_sql_protocol_5_0_6. Historical driver files remain unchanged in rust_values; their retired-API compiler diagnostics remain historical harness evidence, not runtime defects.

The baseline observer manifest from the previous checkpoint was restored byte exactly and its existing source replay passes. New observers and new corpora do not overwrite those earlier observations.

Every historical input is retained unchanged. All 3,473 historical cases now produce current-constructor rejection because they lack required purpose or mutation reason. Their new tagged intent_error values contain the actual source constructor's Display result; C++ uses the corresponding fallible constructor. This is a declared observer response envelope, not a SQL executor error or proof that rejected operations executed. Explicit query_log_intent preflight calls retain their historical ordering and may perform schema/compiler lookups, but rejected query requests never open their query transport.

The C++ query, stream and default-SPI fixtures no longer supply an invented fixed purpose or mutation reason. Valid input controls are separate copies with explicit declarations. Mutation controls supply reasons to each typed child request; this raw observer tree is not the graph-save descendant-reason contract. Additional controls check null, blank, Unicode-whitespace and wrong-type intent input, plus missing/blank mutation reasons in the default-SPI fixture. Runtime intent validation is active with diagnostic output disabled.

Matrix Historical inputs Explicit-intent controls Negative controls Total
Executor defaults 144 144 8 296
Query/cache/diagnostics 381 381 5 767
Stream/drop/diagnostics 880 880 5 1,765
Mutation execution 1,012 1,012 5 2,029
Transaction execution 1,012 1,012 5 2,029
Executor backend capture 44 44 5 93
Actual SQLite cached groups 0 8 0 8
Total 3,473 3,481 33 6,987

The valid controls compare actual original executor calls, transport observations, counter transitions, pending/drop/error behavior, diagnostics, mutation readback and transaction paths. They are not substitutes for the unchanged rejected inputs. Ordinary JSON outputs are exact; there is no output normalization.

The execution observer reuses the existing byte-preserved 5.0.6 original executor observer. Only its input-construction gate is added. Query/stream/default-SPI adapters preserve their historical fixture provider mechanics while replacing retired constructors. No algorithm from the runtime is copied into a provider. See the replay record.

HAVING parameters in cached partition queries

The previous compiler repair retained GROUP BY/HAVING in ranked SQL. These new valid-request controls exposed a second early return in collect_select_bindings. The SQL and policy vectors had four slots, but cached execution supplied only three values, dropping HAVING. The 12 initial differences preserve missing values in both transports and retained metadata.

The collector now visits HAVING before the partition return, matching the original executor. Window-order bindings remain before WHERE/HAVING; ordinary-order values remain after them. The unit regression runs the same statement twice with changed sensitive HAVING text, requires four transport/diagnostic bindings, retains the Masked policy and verifies one compilation for two executions.

The actual SQLite observer uses the original SqlDataServiceExecutor, SqliteDialect and SqliteMutationExecutor with a real schema provider. Eight scenarios each run three statements through one cache, changing the HAVING threshold without changing the SQL shape. All 24 actual database query results, parameter values, parameter policies and parameterized SQL match C++. This covers the real binding path rather than only a recording transport. Native connection/provider/cache objects belong to UserContext; the default PMR resource is disabled and Users are actually deleted.

Terminal verification and remaining integration

Native, canonical modules and targeted ASan/UBSan each pass 8/8. All 6,987 source observations reproduce, including through module consumers. The existing explicit 5.0.6 execution regression passes 1/1 and the previous 7,040-outcome SQL family replay still matches. Module verification covers 465 canonical consumer/fixture bodies and 1,479 exports; pinned format/check covers 1,354 handwritten files. Sanitizer leak scanning is disabled.

The terminal record retains each scope. Six recorded historical-corpus failures now have selected successful follow-ups, leaving 37 unresolved failures from the original full integration. The fresh static inventory does not establish complete-suite acceptance. No fresh full native, MIPS/QEMU or physical-device run is claimed. Producer sources and default generation guards are unchanged. Broad model/configuration, policy/privacy, provider composition, lifecycle and phase exit criteria remain separate requirements.

set -e
source scripts/toolchains/activate.sh
python3 scripts/verify_sql_protocol_5_0_6.py
ctest --preset native --output-on-failure \
  -R '^(executor_defaults_scope_rust_parity|sql_(query_rust_parity|stream_rust_parity|execution_rust_parity|transaction_rust_parity|executor_kind_rust_parity|cached_partition_rust_parity|query_executor_contracts))$'
python3 scripts/verify_sql_protocol_5_0_6.py \
  --cpp-prefix teaql_module_ --build-dir build_modules_native/modules
ctest --preset modules_native --output-on-failure \
  -R '^module_(executor_defaults_scope_rust_parity|sql_(query_rust_parity|stream_rust_parity|execution_rust_parity|transaction_rust_parity|executor_kind_rust_parity|cached_partition_rust_parity|query_executor_contracts))$'
ctest --preset sanitize --output-on-failure \
  -R '^(executor_defaults_scope_rust_parity|sql_(query_rust_parity|stream_rust_parity|execution_rust_parity|transaction_rust_parity|executor_kind_rust_parity|cached_partition_rust_parity|query_executor_contracts))$'

Build those selected C++ targets before replay. The verifier builds current source observers with offline locked dependencies. --record stages separate corpus files only after every declared comparison succeeds. The original historical adapters retain their retired syntax; use the new observer workspace for current execution.