Provisional: extensible random generators and splitting for Gecode 7 (breaking changes)

[zayenz]

Provisional draft — future breaking-change release only

Add modern, user-extensible random number generators with reproducible stream splitting. Splittable SplitMix replaces the old default; xorshift64* provides a smaller-state alternative. The build selects the default engine, while users can supply their own engines through the generic branching APIs.

This is a provisional Gecode 7 design, intended only for a future release that explicitly permits breaking changes. It is not ready for merge and must not ship in Gecode 6.x or any compatibility-preserving release. The default engine, APIs, and performance tradeoffs remain under review.

Generator and search behavior

• Provide indexed splitting: each alternative derives a distinct successor state from the recorded parent state. Recomputing that choice and alternative reproduces the state exactly.
• Support complete-state save/restore independently of seed expansion, including engines with more than 64 bits of state.
• Accept checked 64-bit seeds and complete state on the command line. Test failures and exceptions report exact iteration-state replay commands.
• Store each generator directly in its consumer. Rnd contains exactly 16 bytes with SplitMix or 8 with xorshift64*, without allocation. Copying produces independent state.
• Record only the active randomized brancher's selector states, inline in its choice allocation. Cloning does not split; deterministic branches leave unrelated consumers untouched. Model and custom-brancher transitions remain explicit.
• Keep Space and base Choice unchanged from main, with no extra storage or archive words for nonrandom choices.

Source and binary compatibility, seeded sequences, bounded-draw consumption, random search trees, and randomized choice archives change. No legacy sequence mode is provided. Version and ABI-number changes remain release preparation.

The custom-engine example, engine contracts, migration guidance, and measurements are in examples/random-engine.cpp, docs/random.md, and plans/random.md.

Measured tradeoffs

Local arm64 macOS Release measurements against main (6b7de57), five measured repetitions after warmup:

• One-selector random position/value choice: 48 bytes with SplitMix or 40 with xorshift64*, including state. Nonrandom choice: unchanged at 24 bytes.
• Random archive: 28 or 20 bytes. Nonrandom archive: unchanged at 12 bytes.
• Controlled random-tree time relative to main: SplitMix 1.02x with cloning and 1.16x with recomputation; xorshift64* 1.63x and 2.88x.
• Indexed binary split-plus-draw: about 15.7 ns for SplitMix versus 237 ns for xorshift64*.
• Queens: approximately main's time with SplitMix, 1.06x with xorshift64*. Node counts differ slightly.

These are local measurements, not general performance guarantees. RNG replay along a recorded path does not guarantee parallel scheduling or solution order.

Verification and cleanup

• All five CTests pass with both configured engines, including fault injection, CLI validation, and actual failure/exception state replay.
• All four CTests pass in the reduced static/no-thread build.
• Additional LDSB and Random tests pass with both engines. Earlier integration checks also cover Boolean, integer, set, float, assignment, filtered ties, and FlatZinc restarts.
• Focused tests cover independent copies, untouched model/later-selector state, archived replay, perturbed destination state, sibling alternatives, recomputation, parallel solution agreement, and randomized LDSB archives.
• Custom-engine full-state replay and example output agree across both full builds and the reduced static build.
• Replay CTests now use the existing build fixture, including when selected individually.
• Remove the empty RNG translation unit from both build systems and avoid constructing a temporary choice in LDSB before constructing its final choice.
• Changelog and documentation lead with the generator and splitting features; stale implementation-history and status text is removed.

Other platforms' CI has not been verified locally. Passing these checks does not change the provisional draft status.