Post-rebase review fixes and verification

The combined branch was rebased onto main 39d33bc5d with no conflicts. git range-diff confirmed all four patches were unchanged. The resulting head was cab0cb0c7; its feature commits were 026667c14, d6f51e055 and 075f36a40. Fresh standalone and distributed feature runs each passed 145/145 scenarios. The older integration record retains its original commit IDs, binary hashes and counts; those observations were not relabeled.

Review fixes

An accepted integer-first decimal IN or NOT IN selected the integer kernel and panicked while reading a decimal column. Comparisons now choose the exact decimal kernel if any operand is decimal, preserving each operand’s scale. The existing NULL and short-circuit membership behavior is unchanged.

Wildcard limit errors now identify work, expanded terms, allocation/automaton bytes, pattern count/length or candidate count. The limits remain unchanged. Wrapped errors retain the native sentinel and remote ResourceExhausted mapping.

The QG2 and QG3 reproduction scripts resolve module and build caches through go env GOMODCACHE and go env GOCACHE, including quoted paths with spaces. The benchmark workload and historical timing evidence remain unchanged.

Regression evidence

Expression-execution tests failed before the fix and passed afterward. They cover integer literals/columns, decimal operand order, a decimal later in the list, different scales, NULL candidates and later matching values, for both IN and NOT IN. They also assert query memory is released on closure.

The generator adds 16 scenarios over the existing six-row fixture. Its sum is 12, so 12 equals the scale-0 sum and the scale-2 sum of v * 100, but not the scale-1 sum 1.2. Independent worked expectations cover the full, null-only and empty windows. All 48 bindings run with normal and bypassed plan caches, producing 96 new QL RPCs in each topology; no raw-query comparison was added.

The new qg3_decimal_membership_integer_first_in case was run through the real standalone QL RPC with the old comparator supplied by a Go source overlay. It failed with the expected index-out-of-range panic and Internal response. With the fix, the feature cohort passed 161/161 scenarios in both standalone and distributed deployments, including all 16 new scenarios and their cleanup assertions. Containers ran sequentially with four CPUs, 8 GiB memory and no extra swap, using Go 1.26.10 and -tags=querymetrics.

Machine-readable evidence records the pre-fix failure, pre-fix post-rebase runs, fixed runs, binary hashes and changed source hashes. Fixed-run hashes describe the working tree on cab0cb0c7, not an already committed revision.

Other checks

The QL, parser, decimal, generator and native index package suites passed. Native receive tests verify that the expanded error messages still map to ResourceExhausted and worker failures still fail the query. Full make lint, make check-ql-test-cases and make license-check passed. Both reproduction scripts passed zsh syntax checking and 26 mocked Docker invocations using cache paths with spaces.

No new performance claim is made: comparator selection happens during operator compilation, error formatting occurs only on rejected requests, and the script changes only affect cache-path discovery. The existing benchmark workloads were not rerun or expanded for these fixes.