A failing set_up_before_script produced three different outcomes
depending on incidental details: attributed failure (plain failing
command), silently ignored (failing `cmd && var=x` guard on Bash 3.2 —
execute_file_hook read only the ERR-trap global and discarded the
hook's real exit status), or silent failures with an off-by-one count
(Bash >= 4, where the ERR trap re-fired in the executor's own scope
and returned before record_file_hook_failure ran).

- capture the hook's real status from $? and fall back to the trap
  global; guard the trap's return by FUNCNAME so the executor-scope
  re-fire records instead of silently bailing (both hook executors)
- reset the runner loop to set +euo after sourcing each file: a test
  file's top-level `set -euo pipefail` leaked into the loop and let a
  failing hook abort the whole run mid-suite with no summary
- count hook-failure casualties from the file's own function list and
  run the #829 function cleanup on the failure path too — previously
  leftover functions from earlier files inflated the totals
- harden two src helpers that broke callers under strict mode:
  is_diff_enabled read an unset var, cleanup_testcase_temp_files
  returned 1 on its skip path
- make 13 latently strict-broken tests capture nested failures with
  `|| code=$?` — the loop reset makes --strict honest and deterministic,
  which exposed them (they already failed standalone on main)
- report_json: a jq-less box now skips instead of failing the hook

Closes #836

Claude-Session: https://claude.ai/code/session_01JSeB8UzLCpqst55hAK6dED