[zayenz]

The TSPTW example ran very slowly under the experimental A* and CABS engines: on a generated 60-node instance CABS explored about 100 nodes/s, while BAB managed about 80,000.

Root cause

The engines are fine; the problem is the model. The travel time constraint element(d, x[i], n, x[i-1], n, c[i]) is posted as an index-pair variable over all n² pairs plus an Element::Int over the flattened matrix. Its clone constructor drops the index/value lists, so the first propagation in every clone rebuilds and sorts them. Fixing one customer removes a value from every other x[j] through the channel, which wakes all ~n of these propagators and makes each rescan ~n² entries. A* and CABS clone and propagate every child of every node they expand near the root, so they pay this cost on every node, while BAB mostly works deep in the tree where domains are small. At n=60, CABS spent 2.97 s of a 3 s run in status() on children at depth 1–6, against 17 ms spent cloning.

Change

• examples/tsptw.cpp: replace the matrix element with one TupleSet over (x[i-1], x[i], c[i]), shared by all positions, and post it with extensional.
• gecode/search.hh: document in HeuristicGuidance that a node costs a clone plus propagation at shallow depth, so models should use propagators that stay cheap there. Also document that without a dominance_key, every node offered to the beam is compared with every node in it.

No engine changes.

Measurements

Built-in instances, complete search, median of 20 runs (-mode time -samples 20), RelWithDebInfo:

Instance	Engine	Optimum (main / PR)	Nodes (main / PR)	Time (main / PR)
0 (rbg010a)	BAB	671 / 671	3911 / 3899	32.6 / 20.9 ms
0	A*	671 / 671	3911 / 3883	25.7 / 16.4 ms
0	CABS	671 / 671	2564 / 2608	28.9 / 16.5 ms
1 (n20w20.001)	BAB	378 / 378	527 / 522	48.2 / 23.0 ms
1	A*	378 / 378	527 / 522	40.5 / 21.1 ms
1	CABS	378 / 378	923 / 905	83.0 / 40.1 ms

Generated instances with n customers, nodes explored in 3 s. These were measured with a local file-reader hack that is not part of this PR:

Engine	n=20 (main / PR)	n=40 (main / PR)	n=60 (main / PR)
BAB	208k / 340k	180k / 227k	245k / 292k
A*	24.1k / 159k	1.7k / 55.9k	278 / 23.5k
CABS	39.2k / 183k	3.0k / 79.0k	301 / 35.0k

Validation

• Built out of tree with -DGECODE_ENABLE_EXPERIMENTAL=ON, next to a build of origin/main.
• BAB, A* and CABS find the same optimum on both built-in instances as on main (671, 378), and each search runs to completion.
• gecode-test -test ^Search: 25815 passed, 0 failed.
• git diff --check is clean.