Add recursively productive imperative Mesopy synthesizer

[sileod]

Replaces the initial finite phenomenon composer in this PR with a task-agnostic, recursively productive Python AST synthesizer under reasoning_core/resources.

Design:

• recursive expression synthesis (binary, ternary, safe indexing, helper calls)
• recursive statement/control-flow synthesis (if, for, while) under explicit depth budgets
• helper call graphs whose depth scales with complexity
• generated-program recursion with a decreasing ranking argument for runnable programs
• input-controlled RecursionError using the same recursive helper for failing runnability cases
• explicit multidimensional complexity control via MesopyComplexity: statements, expression depth, control depth, functions, call depth, dataflow depth, loop bound
• structural features (ast_nodes, AST depth, control depth, call depth, loops, branches, calls, dataflow depth) for profiling/complexity tasks
• opt-in dynamic line-step profiling so ordinary generation does not pay tracing overhead
• all controlled-code-execution phenomena remain requestable, plus recursion/comprehension/mapping/helper/control-flow phenomena

Runnability anti-shortcut behavior:

• paired samples use identical source with one successful and one failing call
• supports IndexError, ZeroDivisionError, ValueError, KeyError, and RecursionError
• the same hazard families also occur in successful execution samples via safe_hazard_rate, so suspicious syntax is not itself a label

Productivity tests compare complexity levels 0, 3, and 6 and require increasing median AST size plus increased AST/control/call/dataflow depth. The generator is therefore not just concatenating from a finite segment catalogue; expression and control-flow constructors recurse as budget increases.

Local stress/benchmark results on this implementation:

• 1,000 executable samples: no generation failure
• 500 paired-runnability samples: no generation failure
• default executable generation: ~245 samples/s before the extra semantic relevance experiment; ~200+/s in the stricter local variant
• default paired runnability: ~178 pairs/s before the extra semantic relevance experiment; ~130+/s in the stricter local variant
• complexity level 6 executable generation: ~50 samples/s before the extra semantic relevance experiment
• targeted regression suite: 7 passed locally

CI tests use deliberately loose performance limits (100 executable samples < 3 s, 50 paired samples < 3 s) to detect regressions without encoding machine-specific benchmark numbers.

This PR intentionally only adds the reusable resource and its tests. Migrating controlled_code_execution, code_execution, runnability, and profiling tasks can be reviewed separately against this common generator distribution.