Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@cpp/src/branch_and_bound/branch_and_bound.cpp`:
- Around line 743-745: The logged/exported counter was switched to
exploration_stats_.total_nodes_explored but solve_node_deterministic() only
increments exploration_stats_.nodes_explored, so deterministic runs report zero;
update the producer-side accounting inside solve_node_deterministic() (or the
deterministic node-visit path) to also increment
exploration_stats_.total_nodes_explored (or increment both
exploration_stats_.nodes_explored and exploration_stats_.total_nodes_explored)
wherever nodes are counted, ensuring the same counter used by the log/return
paths is updated for deterministic and opportunistic solves.
- Around line 1563-1572: The debug call in branch_and_bound.cpp using
settings_.log.debug mixes printf-style format specifiers with a "{}" placeholder
so nodes_since_last_check is never printed; change the "{}" to the appropriate
printf specifier (e.g., "%d") so the format string matches the arguments
(num_nodes, current_progress, current_abs_gap, nodes_since_last_check,
progress_since_last_check, gap_reduction, tree_size_estimate) and ensure the
order of placeholders aligns with those variables.
- Around line 2982-2985: The restart checkpoint mixes internal and user-space
units: set exploration_stats_.restart_gap_at_last_check using user-space units
(call compute_user_abs_gap() or otherwise convert upper_bound_ -
get_lower_bound() into user-space) so it matches what should_restart() compares;
update the initialization in the same block (replace upper_bound_ -
get_lower_bound() with compute_user_abs_gap()) to keep units consistent with
should_restart() and compute_user_abs_gap().

In `@cpp/src/branch_and_bound/worker.hpp`:
- Around line 230-237: bfs_worker_t::reset_state() currently clears node_queue
and zeroes total_active_diving_workers but does not reset the per-strategy
active_diving_workers array, leaving stale counts; modify reset_state() to also
zero all entries of active_diving_workers (e.g., fill or loop over
active_diving_workers to set each element to 0) so per-strategy state is
consistent after restart, while keeping the existing resets for
total_max_diving_workers, total_active_diving_workers, is_active, and
lower_bound.

---

Outside diff comments:
In `@cpp/src/mip_heuristics/diversity/lns/rins.cu`:
- Around line 217-232: The child OpenMP task binds fj_cpu->preemption_flag to
the stack-local fj_halt_flag which can dangle if an exception is thrown before
the task completes; fix by giving the task a stable, RAII-managed flag and
ownership so it is valid for the task's lifetime: allocate the halt flag with
shared lifetime (e.g., a heap/shared_ptr<atomic<bool>> or embed it inside a RAII
task object) and set fj_cpu->preemption_flag to that stable address, ensure the
RAII object owns joining/stopping the task and is destroyed on all exit paths;
update the code paths around fj_cpu, fj_halt_flag, and cpufj_solve to use that
RAII-managed flag so preemption_flag never points to stack memory.