Nothing runs test/extension_upgrade.sh: it has pg_upgrade's exemption without pg_upgrade's invocation

[jdatcmd]

Summary

test/extension_upgrade.sh landed in #389 and is a good guard: it installs a released
build, creates a columnar table with rows, installs the tree under test over it, runs
ALTER EXTENSION UPDATE, and then requires reads, writes, table creation, a
maintenance call, and that every C function's link name is namespaced.

Nothing runs it.

The gap, precisely

It is listed in harness_selftest.sh's not_a_suite:

lib|portlib|run_all_versions|...|pg_upgrade|extension_upgrade) return 0 ;;

That exemption is correct -- it needs two builds and a restart, which the ordinary
suite contract cannot express, so it must not be in SUITES.

But look at what pg_upgrade gets in exchange. It is exempt from the registration
check and explicitly invoked by the runner:

test/run_all_versions.sh:505:   if bash "$SRCDIR/test/pg_upgrade.sh" "$_old" "$_new" "$_mode" ...

That invocation is what closed #257, titled "Wire test/pg_upgrade.sh into an opt-in
gate (not per-PR) so #256's coverage doesn't rot".

extension_upgrade.sh has the exemption and not the invocation. So it is in
neither the matrix, nor CI, nor the Makefile:

$ grep -rn "extension_upgrade" .github/workflows/ Makefile
(nothing)

It runs only when a human types its name. That is the same gap #257 existed to close,
one suite over, and it is worth more here than there: the guard exists precisely
because the failure it catches is invisible until a user upgrades.

Second, smaller problem: it cannot run in the documented dev loop

The container loop copies the tree without .git (documented in
docs/testing.md), and the suite needs tags to find the old build:

FAIL  v1.0-alpha is not present. Fetch tags, or pass an explicit ref:
      git fetch --tags && test/extension_upgrade.sh <pg_config> <ref>

Reviewing #389 I had to copy .git into the container to exercise it at all. So even
a person who remembers to run it hits a wall in the standard environment.

What would fix it

1. Invoke it from run_all_versions.sh the way pg_upgrade.sh is invoked, on the
   same opt-in terms, so the matrix covers it.
2. Either let it take a non-git source for the old build (a tarball or an installed
   prefix), or state in its header that it needs a real checkout with tags, so the
   failure reads as a precondition rather than a bug.

Neither blocked #389, which is merged and green: full matrix PG18 + PG19 ALL VERSIONS PASSED, and I verified the upgrade path by hand. This is about the guard surviving
the next six months.

[ChronicallyJD]

Plan: #400, which is open and green

Fixed in #400, with your second round folded in.

First push wired it into run_all_versions.sh under PGC_RUN_UPGRADE=1 and added the
directory form. You then ran it in the container and it went red, because the runner passed
no second argument and so always took the ref form, in a tree with no .git. The directory
form was reachable by a person and not by the gate.

Second push does both things you suggested:

• PGC_UPGRADE_OLD_SRC names the old source and the runner appends it when set
• absent any named old source, a tree with no .git reports SKIP and exits 2, and the
  runner records SKIP in the summary, following PGC_SKIP_TIMING

The rule I settled on for skip versus fail is who asked for what. Name an old source
that cannot be honoured and it is a failure. Name nothing, in a tree that cannot supply one,
and it is a skip. The skip is loud, prints its reason, and is never reported as a pass,
because a gate reporting green having run nothing is the defect this suite exists for.

Verified as exit codes rather than output text, since the runner keys on the code:

no argument, no .git       -> 2  SKIP
explicit ref, no .git      -> 1  FAIL
explicit directory         -> 0  PASS
PGC_UPGRADE_OLD_SRC, no .git -> 0  PASS
env var naming a bad dir   -> 1  FAIL

That last row found a real defect while checking: a path that did not exist fell through to
the ref branch, so a question about a directory got an error about git checkouts. Fixed to
say what is actually wrong.

Your container scenario reproduced, not approximated: no .git anywhere, through the
runner, SKIP without the variable and PASS with it.

Full gate on the final tree: ALL VERSIONS PASSED, with extension_upgrade PG18 in the
summary beside both pg_upgrade modes.

Closes when #400 merges.