fix(tests): e2e.sh discarded all mix test output, making CI undebuggable

[hyperpolymath]

The problem

tests/e2e.sh:118 judged the whole Elixir suite with:

mix test --trace 2>&1 | tail -5 | grep -q "0 failures"

The pipe throws the entire run away. When the suite fails, CI prints exactly this:

  FAIL: Elixir unit tests
  Results: PASS=7  FAIL=1  SKIP=0

No failing test name. No assertion. No stacktrace. E2E — Elixir Scanner Pipeline has been
red on main with nothing whatsoever to diagnose from — the job reports that it failed
while actively destroying why.

Changes

Capture the run to a log, judge it by mix's own exit status, and echo the tail on
failure. Exit status is also strictly better than grep "0 failures", which mis-reports a
crash that never reaches the summary line as a pass-check failure indistinguishable from a
test failure.

Verification

Stubbed a failing mix and confirmed the diagnostic now surfaces:

  FAIL: Elixir unit tests
--- mix test output (tail 120) ---
  1) test foo (Hypatia.ReflexiveTest)
     ** (exit) no process
--- end mix test output ---

bash -n clean; shellcheck reports nothing on the new block.

What it immediately revealed

Reproduced locally (Elixir 1.18.3 / OTP 27):

Run	Result
mix test (full suite)	1398 tests, 20 failures
mix test test/reflexive_test.exs (alone)	16 tests, 0 failures

All 20 failures are in test/reflexive_test.exs, all the same shape:

** (exit) exited in: GenServer.call(Hypatia.Supervisor, {:terminate_child, Hypatia.SelfDiagnostics}, :infinity)
    ** (EXIT) no process: the process is not alive ... possibly because its application isn't started
  test/reflexive_test.exs:29: Hypatia.ReflexiveTest.take_supervised/1

Since the file passes in isolation, Hypatia.Supervisor is being torn down by something
earlier in the full run — a test-isolation defect, not a product bug. take_supervised/1
(test/reflexive_test.exs:29) does terminate_child + delete_child and restores via
on_exit; under :one_for_one a child crash-looping past max_restarts would also take the
whole supervisor down.

That is a separate fix and is not attempted here — this PR only stops the harness from
hiding it. Filed separately.

Scope

One file. Does not change what is tested, only whether a failure is legible.

🤖 Generated with Claude Code