refactor: the six columnar.h declarations nothing outside their file uses (#496)

[ChronicallyJD]

A first, deliberately small bite of #496: the declarations in columnar.h that nothing outside their defining file uses. The issue counts 10 of these and calls them the cheapest thing available. Six of them are, and this is those six.

No behaviour changes. 93 lines removed, 4 added.

What changed

symbol	file	change
PgColumnarCodecAvailable	columnar_compression.c	deleted — no caller anywhere, including its own file
PgColumnarReadBloomList	columnar_metadata.c	deleted — same
PgColumnarComputeFullyDeletedGroups	columnar_metadata.c	static
PgColumnarVMSetVisible	columnar_visibilitymap.c	static
PgColumnarVMClearVisible	columnar_visibilitymap.c	static
PgColumnarVMIsVisible	columnar_visibilitymap.c	static

The three VM helpers are the low-level primitives; columnar_tableam.c uses the wrappers PgColumnarVMClearForRow and PgColumnarVMSetVisibleForRelation, which stay exported. Deleted rather than made static where there was no in-file caller either, because an uncalled static is an unused-function warning, and this matrix treats warnings as failures.

Why it is six and not ten

Four of the ten cannot be touched, and the reason is one the compiler will not give you. PgColumnarBlockNextRun, PgColumnarBlockReaderInit, PgColumnarCheckFreeSpaceNoOverlap and PgColumnarEncodingName are named in test/ or sql/ — they are reached through the fmgr by name, from CREATE FUNCTION ... AS 'pgcolumnar', 'sym', not by a C caller.

Making one of those static compiles and links cleanly and then fails at run time when the SQL binding cannot resolve the symbol. So "no .c file references it" is not sufficient evidence that a declaration is private, and the audit that produced this list disqualifies anything with PG_FUNCTION_INFO_V1 or a mention under test/ or sql/ before proposing it.

That distinction is the part of this worth keeping for the rest of #496.

Verification

A successful build proves less than it looks like here: a shared library links with undefined symbols permitted, so a stale reference to a now-static function would not fail make — it would fail at dlopen, i.e. at CREATE EXTENSION. So the load-bearing evidence is the suites, not the compile.

• Five majors (15.18, 16.14, 17.6, 18.4, 19beta2): build clean, 0 warnings each.
• index_only 27 checks / 0 fail — the suite that drives the visibility-map path these three helpers implement.
• Full matrix on assert builds of PG18 and PG19: ALL VERSIONS PASSED (132 and 134 suites). Every one of those does CREATE EXTENSION, which is where a missing symbol would surface.

I also tried to verify linkage directly with nm and could not make it say anything trustworthy: the .so exports nothing in the dynamic table (194 symbols exist only in the full one), and the still-shared wrappers report the same symbol type as the newly-static ones, so the check could not distinguish the two cases it was supposed to distinguish. Recording that rather than quoting the output it did produce.

The gate ran one commit before #524 merged; that commit adds bench/provision.sh only, and no suite executes or reads it (bench_guards inspects cb_guards.sh), so the matrix outcome cannot turn on it.

Refs #496.

[jdatcmd]

Approved. +4/-93, and the four files it touches are the four that own the declarations it removes.

Verified each one is genuinely single-consumer rather than trusting the count, and my first attempt at that check was wrong in an instructive way. A plain grep -rl PgColumnarVMSetVisible reports two files, which would make static a compile error. It is a prefix of PgColumnarVMSetVisibleForRelation. With grep -rlw:

PgColumnarVMSetVisible            src/columnar_visibilitymap.c
PgColumnarVMClearVisible          src/columnar_visibilitymap.c
PgColumnarVMIsVisible             src/columnar_visibilitymap.c
PgColumnarCodecAvailable          src/columnar_compression.c
PgColumnarComputeFullyDeletedGroups  src/columnar_metadata.c
PgColumnarReadBloomList           src/columnar_metadata.c

Six for six. CI agreeing is the stronger evidence — static on a symbol used from another translation unit does not link — but the word-boundary check is what makes the claim readable rather than inferred from a green build.

Worth recording that my sloppy probe matched a longer name, since that is the same failure shape we have been cataloguing all day pointed at a review rather than at code. A prefix match is an instrument that cannot distinguish two symbols.

On the change

This is the right first slice of #496 and the right size. Six declarations that nothing outside their own file uses, made static, with the header shrinking by the same amount. No behaviour, no reordering, nothing that needs a matrix argument beyond the one CI already gives.

The value is not the 93 lines. It is that columnar.h stops advertising six things as an interface when they are implementation detail, so the next reader looking for the actual seam has six fewer false leads. #496's complaint is that 65% of the header's declarations have a single consumer; this is a demonstration that the number is real and that shrinking it is mechanical rather than risky.

Merging.