ci: make the FreeBSD job actually build, and fail when it does not

[jcelerier]

The FreeBSD job has been green since it was added, and it has never compiled score even once.

What is broken

ci/freebsd.build.sh started with:

#!/usr/bin/env bash -e

env(1) does not split its first operand, so this asks for a program literally named bash -e. On FreeBSD that is:

env: bash -e: No such file or directory

and the script exits 127 without running a single line. It is not a FreeBSD quirk either — GNU env says the same thing (use -[v]S to pass options in shebang lines). #!/bin/bash -e, which every other ci/*.build.sh uses, is not an option here: on FreeBSD bash is /usr/local/bin/bash.

Here is the whole build step of run 31948475911 on master:

13:06:10.1976660Z =============== FreeBSD Deps Success =====================
13:06:10.1977851Z =============== FreeBSD Start Build =====================
13:06:10.2000273Z env: bash -e: No such file or directory
13:06:10.2001486Z =============== FreeBSD End Build =====================

Two milliseconds, and a green checkmark.

Why CI did not notice

.github/workflows/bsd.yml ran the script as:

./ci/freebsd.build.sh && echo "=============== FreeBSD Build Success ====================="
echo "=============== FreeBSD End Build ====================="

The && turns the failure into a skipped echo, and the step's exit status is that of the last command in the block — the unconditional End Build echo, which always succeeds. So no matter what the script does, the step is green.

Adding set -e to the block is not enough on its own: errexit is specified to ignore any command of an AND-OR list other than the last one, so false && echo hi does not stop an errexit shell. The && has to go.

The change

• ci/freebsd.build.sh: keep env (bash is not in /bin there), move the option into the script as set -e, with a comment so it does not get folded back into the shebang.
• .github/workflows/bsd.yml: set -e in the run block, and run the build script as a plain command instead of the left-hand side of an &&.

The deps line is deliberately left alone — #2202 rewrites it along with ci/common.deps.sh, and its version of this hunk is a superset of this one, so the two converge.

How it was verified

In a FreeBSD 15.0 VM built from the same vmactions/freebsd-builder image the CI uses, with the packages this workflow installs. The workflow's build block was replayed against a cmake stub that fails on --build:

	script	step exit status
master, build fails	env: bash -e: No such file or directory	0 (green)
this PR, build fails	FAKE CMAKE: build FAILED	1 (red)
this PR, build succeeds	Build Success	0 (green)

and running master's script by hand on FreeBSD gives exit status 127.

Then the real thing: ./ci/freebsd.build.sh on a full checkout with all the add-ons, which now runs and reports what it finds.

What it finds

Once the script actually runs, master does not compile on FreeBSD. One add-on is responsible: score-addon-sysinfo vendors lfreist/hwinfo, whose "unix" backend is really the Linux one — it needs <netpacket/packet.h> and _SC_AVPHYS_PAGES, neither of which FreeBSD has, and it reads /proc and /sys throughout. That is fixed in ossia/score-addon-sysinfo#3, which skips the add-on on the BSDs and explains why a real BSD backend belongs upstream rather than in a shim.

Merge that one first. ci/common.deps.sh clones the add-on from its default branch, so as soon as it lands the FreeBSD job here goes green on its own — and with this PR, green will finally mean something. Merged the other way round, this PR turns the job red until the add-on fix follows.

With the add-on out of the way, the whole of master compiles and links on FreeBSD 15.0.

[jcelerier]

Verification transcript, FreeBSD 15.0 VM (same vmactions/freebsd-builder image as the job, same package set).

1. master, replayed with a cmake stub that fails on --build

############ BEFORE: master shebang + master workflow block, build fails ############
=============== FreeBSD Start Build =====================
env: bash -e: No such file or directory
=============== FreeBSD End Build =====================
>>> workflow step exit status = 0

############ AFTER: fixed shebang + fixed workflow block, build fails ############
=============== FreeBSD Start Build =====================
FAKE CMAKE: configure ok
FAKE CMAKE: build FAILED
>>> workflow step exit status = 1

############ AFTER, build succeeds ############
=============== FreeBSD Build Success =====================
=============== FreeBSD End Build =====================
>>> workflow step exit status = 0

############ master shebang alone (script never runs at all) ############
>>> script exit status = 127

2. the real build, full checkout with every add-on

With score-addon-sysinfo as it is on its main:

[43/44] Building CXX object .../hwinfo/src/linux/network.cpp.o
FAILED: [code=1] .../hwinfo/src/linux/network.cpp.o
  fatal error: 'netpacket/packet.h' file not found
[44/44] Building CXX object .../hwinfo/src/linux/ram.cpp.o
FAILED: [code=1] .../hwinfo/src/linux/ram.cpp.o
  error: use of undeclared identifier '_SC_AVPHYS_PAGES'
ninja: build stopped: subcommand failed.

With ossia/score-addon-sysinfo#3 applied:

 -- score: configuring score-addon-sysinfo
-- score-addon-sysinfo: hwinfo has no FreeBSD backend, skipping
...
########## STEP 3 script exit status = 0 ##########
-- Installing: /build/install/bin/ossia-score

So the FreeBSD job on this PR is expected to go red until ossia/score-addon-sysinfo#3 lands — that red is the fix working. ci/common.deps.sh clones the add-on from its default branch, so nothing further is needed here once it is merged.