fix(storage): probe before the pipe, not inside it (#462)

[fujibee]

Declared reviewers: 1

Change class: performance, no behaviour change. Every statement runs as
before; some of them stop starting a second sqlite3 process to ask a question
that was already answered.

What is wrong

agmsg_sqlite memoises the escape-flag probe so it costs one sqlite3 process
per shell rather than one per call — that is #462, and the reason is written
above the function.

The right-hand side of a pipeline is a subshell. It inherits a memo, but a
memo it sets there dies with it. So a process whose first database access is
piped records nothing, and every piped call after it probes again.

Measured, in a shell whose first access is piped:

pipe#1 = 2 sqlite3 processes   probed_in_parent=unset
pipe#2 = 2                     probed_in_parent=unset
pipe#3 = 2                     probed_in_parent=unset

Against the same call after an unpiped one has warmed it:

argv (cold) = 2      pipe#1 = 1      pipe#2 = 1

It does not converge: the piped form pays the probe forever.

Where

Ten sites, derived with grep -rn '| agmsg_sqlite' scripts/ rather than
enumerated. Three of them arrived with #895 — my change, which moved
data-sized SQL onto stdin and was reviewed and cleared without the shell rule
underneath it being noticed. The rest predate it, including the migration
helper and the legacy import path.

A redirection is not a pipe. agmsg_sqlite db < file runs in the current
shell and memoises normally, so the four < "$sql_file" sites need nothing —
including the one #899 just added.

The fix

agmsg_sqlite_warm beside agmsg_sqlite, called on the line before a
pipeline. Two of the driver's statements go through a helper that warms, so the
rule has one place there rather than one per statement.

Measured

tests/test_remote_sync.bats     33/33   (1 new)
tests/test_storage.bats         24/24
check-enforced-assertions       638, at the baseline

The new case is a scan, not a list: it derives the piped sites and fails on
any without a warm above it. Written that way because the sites move — three
appeared in a single day, in a change that two people read.

warm removed from the helper       -> red
an unwarmed pipe planted in sqlite.sh -> red, and it names the file and line
none (control)                     -> green

It also asserts the scan can see at least eight sites, so an instrument that
matched nothing could not pass as "everything is warm".