Poll timeout/interrupt/memory-limit inside dive loops; fix pulling should_stop

[legraina]

Summary

Two fixes for pathological solves that can wedge a caller for minutes inside a single solve() call.

1. Poll should_stop() inside dive loops (7ca2273)

A dive is a DFS with backtracking: on pathological instances a single dive can enumerate exponentially many partial paths, while all per-iteration should_stop() checks sit between dives — so timeout_s and the external interrupt never fire. The dive loops in GreedyAlgorithm, TabuSearchAlgorithm, and ImprovingTabuSearch now poll should_stop() every 4096 steps.

Also fixes PullingDominanceAlgorithm::main_loop, which historically only checked the iteration budget: it now uses should_stop(i), which covers the budget plus timeout_s and the external interrupt, like the base DominanceAlgorithm.

2. Poll the memory limit inside dive loops (b5665b5)

Each dive extension allocates a label, so a pathological dive is an RSS runaway as well as a time sink — and only the dominance main loops consulted MemoryLimitHelper. The same 4096-step poll now also checks memory_limit_.is_exceeded(). The memory check stays out of should_stop() itself on purpose: reading process RSS is a real syscall, and should_stop() runs every iteration of the hot dominance loops (which sample memory on their own throttled memory_check_interval cadence).

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR hardens several RCSPP algorithms against pathological “dive” behavior by adding periodic stop-condition polling inside deep inner loops, and fixes PullingDominanceAlgorithm to honor the same stop logic as the base dominance loop.

Changes:

• Add periodic polling (every 4096 steps) for timeout / external interrupt / memory-limit during DFS-style dive loops in tabu-search variants and greedy backtracking.
• Update PullingDominanceAlgorithm::main_loop() to use should_stop(i) instead of only checking the iteration budget.
• Document the pathological long-run / OOM scenarios these guards address (ROADEF setB-01).

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
cpp/rcspp/algorithm/tabu_search.hpp	Adds periodic timeout/interrupt/memory polling inside dive_to_sink() to avoid wedging in a single dive.
cpp/rcspp/algorithm/pulling_dominance_algorithm.hpp	Fixes override loop to use should_stop(i) so timeout/interrupt are respected.
cpp/rcspp/algorithm/improving_tabu_search.hpp	Adds the same periodic polling inside dive_to_sink() for the improving tabu search.
cpp/rcspp/algorithm/greedy.hpp	Adds periodic polling inside greedy backtracking/extension to avoid long uninterruptible runs.

Suppressed comments (1)

cpp/rcspp/algorithm/greedy.hpp:109

• The 4096-step poll is currently only evaluated once per outer loop iteration, but the inner while (extend_label(...)) can perform an arbitrarily long greedy “dive” (many label allocations / extensions) without ever re-checking timeout/interrupt/memory. To actually make greedy dives interruptible, poll per extension attempt (e.g., inside the inner loop or by incrementing steps for each call to extend_label()).

                bool extended = false;
                while (extend_label(path_.back().first)) {
                    extended = true;  // successfully extended
                }
                if (extended) {

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.