perf: skip column decoding for cached entities and precompute constructor metadata

[zantvoort]

Implements the remaining checklist item of #246: skip parent column decoding on cache hits. The mapper's early cache lookup already skips construction when a nested entity is cached, but every column of every row was still decoded out of the JDBC ResultSet before the mapper ran, including the repeated parent columns the lookup then ignored.

Column skipping

A new ColumnSkipper probes the same caches as the mapper's early lookup, but before column extraction. For each entity column region it decodes only the columns up to and including the primary key; on a cache hit the remaining columns are left undecoded, and the mapper resolves the entity from the primary key alone.

• Skip regions (start, end, PK offset, entity type) are computed once per record type alongside the compiled argument plan, including nested regions (a city region inside an owner region). Outer hits skip inner regions wholesale; outer misses give inner regions their own probe.
• Cache policy parity: nested regions consult the transaction-scoped EntityCache at REPEATABLE_READ+ and the query-scoped WeakInterner otherwise; the top-level region only consults the entity cache, since top-level records are never interned. This mirrors RecordStep and newInstance exactly.
• Correctness guards: column access stays in ascending order (safe for drivers that only support forward column access within a row), and interner hits are pinned in a strong reference list until the next row so a garbage collection between probe and mapping cannot clear the weakly held entity.
• ObjectMapper exposes the skipper via a default method returning null, so only the record mapper opts in; sealed-type mappers and the JPA path are untouched.

Constructor metadata

Record construction carried per-invocation reflective overhead on the row mapping hot path: getParameterTypes() and getParameters() clone their arrays on every call, setAccessible(true) ran per call, and the enum step resolved its mapper per value. Constructor metadata (parameter names, non-null flags, primitive flags, the accessible constructor instance) is now precomputed and cached per constructor; enum mappers are resolved once at plan compilation. Error messages and diagnostics are unchanged.

Testing

ColumnSkipperIntegrationTest counts positional column reads through a JDBC proxy and asserts the exact expected number for the pet → owner → city graph: full decode on first occurrence, key-only decode on repeats, full decode for null foreign keys, nested city regions, and interner identity across rows. Full suites green: storm-core 2264, storm-kotlin 1627, storm-java21 517.

Measured effect (#214 suite, PostgreSQL 17, same session)

Workload	Before	After
joinWithMapping10	577 µs	444 µs (−23%)
joinWithMapping100	1281 ± 372 µs	913 ± 69 µs (−29%)
joinWithMapping1000	3636 µs	3325 µs (−9%)
projection	975 µs	884 µs (−9%, constructor metadata)
singleRowById	unchanged	unchanged (no duplicate regions, probe cost not measurable)

The collapsed error bar on joinWithMapping100 (±372 → ±69) reflects the removed allocation pressure.