datastore/.claude/skills/review/references.md

Review references

Detailed procedures for triggered expansions that fire on a small fraction of PRs. Read the relevant section only when its trigger appears in the diff; the base checklist in SKILL.md names the trigger and points here.

Build flag that drops a global side-effect or weakens numeric semantics

Trigger: the diff changes compiler/build flags that remove a documented side-effect or relax FP behavior across many translation units — -fno-math-errno (libm stops setting errno), -ffast-math / -ffp-contract=fast / -fassociative-math / -freciprocal-math / -fno-signed-zeros (reordered/contracted/round-changed FP → non-reproducible results), -fno-trapping-math, strict-aliasing relaxation, and similar.

The PR's implicit promise is "no code relying on the old behavior is affected"; treat that as the contract and verify it. A grep over the affected symbol is not proof, because a match only matters if it is (a) compiled with the flag and (b) actually depends on the specific behavior the flag removes, not merely mentions the symbol.

• (a) build graph (general): a match in code not compiled with the flag cannot be affected — rust_vendor/cargo crates, EXCLUDE_FROM_ALL, disabled ENABLE_* options, or an uncompiled part of a contrib (e.g. a vendored DB's server backend when only its client library is built).
• (b) what a true consumer looks like is flag-specific — work it out per flag, don't reuse one flag's shape for another. For -fno-math-errno the affected behavior is errno after a <math.h> call, so most matches are false: code that sets errno (a libm's own errno = EDOM / with_errno(...), an error-code table, return ERANGE) is not a consumer — only a read after a math call is; and a read whose producer the flag does not touch (strtod/strtol/scanf, a syscall) is unaffected. For reproducibility-affecting FP flags (-ffast-math/contract/associative/signed-zeros) there is no "setter/reader" split at all — the consumer is anything that assumes bit-identical results: cross-platform/cross-ISA reproducibility, distributed-merge/aggregation consistency, and committed test references — so the cost is acknowledged and reasoned about, not grep-audited.

Also check scope intent: flags appended to COMPILER_FLAGS before add_subdirectory(contrib …) propagate into vendored libraries. Contrib coverage may be deliberate (often it is where the speedup lives) or unsafe, but it must be a conscious, stated decision with any exceptions carved out per target. A real consumer that silently loses the behavior is a correctness/compatibility Blocker; an unstated contrib-wide reproducibility change is at least a Major.

Native protocol / native format spec sync

Trigger (protocol): the diff touches the native TCP protocol — src/Core/Protocol.h, src/Core/ProtocolDefines.h, src/Core/Protocol.cpp, packet handling in src/Server/TCPHandler.* or src/Client/Connection.*, the DBMS_TCP_PROTOCOL_VERSION / DBMS_MIN_REVISION_* constants, packet types, handshake/version negotiation, or the wire layout of any non-Block message.

Trigger (format): the diff changes the Native format — its wire/serialization format, type encodings (LowCardinality, Array, Map, Variant, Dynamic, JSON), the block/column structure, the compression frame, NativeReader/NativeWriter, or docs/en/interfaces/formats/Native.md.

Verify the corresponding specification is updated in the same PR: docs/en/interfaces/specs/NativeProtocol.md for protocol changes, docs/en/interfaces/specs/NativeFormat.md for format changes. The spec is the canonical reference third-party native clients (ch-go, clickhouse-go) are built against; letting it drift forces re-deriving the protocol/format from C++ source. This applies to new features, bug fixes, and behavior changes alike. Flag a missing or stale spec section as a Major, naming the packet/version/field/encoding the diff changed.