fix(engine): treat a crashed test/security gate as failed, not green (ISSUE-57)

[ecukalla]

Problem

The Claude-driven test and security gates decide their verdict purely by "did the agent write its failure file?". The only code that synthesized that file when the agent didn't was fl_gate_timeout, which bailed on every non-timeout exit (fl_timed_out "$rc" || return 0).

So any non-timeout, non-zero exit of claude -p — auth/API error, OOM, killed, a rejected --settings, budget truncation, a bad FL_CLAUDE binary — wrote no failure file. The gate then read as a pass, and the run shipped OUTCOME=green / exit 0 with the test and security gates never actually having run. This is the exact failure mode gate_pipeline already guards against (it's fail-closed on rc -ne 0).

Fix

Generalize fl_gate_timeout → fl_gate_crash and change its guard from fl_timed_out "$rc" || return 0 to [ "$rc" -ne 0 ] || return 0, so it synthesizes the failure file on any non-zero gate exit, called unconditionally from _run_gate.

• The timeout path (124/137) keeps its distinct "timed out / exceeded FL_PHASE_TIMEOUT" wording.
• Other non-zero exits get a generic "crashed (exit N)" message.
• rc == 0 stays a no-op — a clean agent run that wrote nothing still reads as a pass.
• An existing failure file the agent already wrote is preserved ([ -f "$ff" ] && return 0).

Testing

New bats test (a gate whose claude call crashes is marked failed, not read as green): stubs claude to exit 3 only on the test-gate prompt without writing a verdict, runs one iteration, and asserts the run ends not-green (exit 1) with a synthesized failures/test.md naming crashed (exit 3).

• bats --print-output-on-failure tests/ — green (via feature-loop's test gate)
• bash -n bin/feature-loop parses clean

Closes #57