refactor: split the backtrack engine into focused collaborators

[rasros]

What changed

Decompose the backtrack BacktrackSolver/engine monolith into focused, single-responsibility pieces. Seven commits, each behavior-preserving:

• Extract the 713-line B&B engine (private inner class ResumableMinimize) out of BacktrackSolver into its own top-level internal class. BacktrackSolver (976 lines) becomes a ~245-line Solver/Optimizer facade; the engine takes the solver via its constructor.
• Extract PhaseSaving (phase-saving + target-phasing + rephase schedule), shared by the satisfaction path (driveSearch) and the engine — removing the duplicated phase logic and the loose array threading through applyPhase/capturePhase/backjumpAndLearn.
• Extract DeadlinePoller (the adaptive cancel-check cadence), also shared by both paths.
• Extract RestartController (Luby/Glucose restart policy + the Luby math), shared by both paths.
• Extract PortfolioBoundExchange (level-0 cross-arm objective-floor and variable-bound sharing).
• Decompose the 269-line runUntilEvent loop body into named descendStep / backtrackStep methods.

The incumbent state (best/bestObj/objVarBest) stays in the engine by design: it is read across pruning, both terminal builders, and the leaf handler, so hiding it behind a collaborator would be a leaky abstraction rather than an improvement.

Why

BacktrackSolver.kt (52K) welded a thin API facade to a giant stateful search engine, and the phase/cancel/restart bookkeeping was duplicated between the satisfaction and optimize paths. Splitting them separates concerns and dedups the two drivers.

Testing

Behavior-preserving throughout; covered by the existing backtrack/optimize suites, full check lintDocs green across JVM + native + wasm. Benchmarked A/B (core suite, branch vs merge-base): identical solve outcomes, and wall-time differences within the run-to-run noise floor (the same binary swings ~230ms between repeats on these sub-second instances), i.e. no measurable regression.