Support bounded rule loops and restore ordered array scans

[zhoubot]

Pure behavioral rules could not express ordered prefix computations without manually expanding each step. This change implements bounded ordinary for name in range(INTEGER_LITERAL) through the existing Python capture and MLIR lowering flow and restores the original array_scans example. No decorator, Table scan method, IR opcode or alternate compiler route is added.

Loop lowering preserves logical Integer induction, existing typed assignment boundaries, final binder values and zero-trip behavior. It rejects unsupported effects and syntax before expansion, prepays source-wide resources, retains iteration/call provenance and rolls back failed staging. Unchanged branch values reuse the existing SSA only after declaration/binding validation. The example retains all four ordered scalar/aggregate scans, complete prefix snapshots and the two original depth-one queues.

Deep-record validation exposed redundant C++ default member initializers. Direct nominal Struct fields now rely on their initialized leaves; scalar, Enum and Table initialization, layout and aggregate construction are preserved. For the same depth-32 final IR and compiler, one measured generated-C++ syntax check fell from 154.77 s to 0.41 s. This is a single workload measurement, not a general scaling claim.

Validation:

• Source interface/ledger/rollback native suite: 73/73.
• Existing API gate: 22/22; installed source-system execution: 43/43.
• Existing local binding, four-state binding, branch, default, scalar-helper and multi-writer regressions: 6/6.
• array_scans: 1,031 Work samples agree across native workers 1/2 and Verilator; 495 accepted tokens, 492 retired and three reset drops. Closed system: 16 cycles / 32 sampling epochs, both backends. Both standalone CTests pass.
• Complete source-import and transformed IR for DUT/bench retained and byte-matched to published units; both backend outputs consume the same verified final artifact.
• Public bounded-loop gate: native workers 1/2 and Icarus six known/X/Z rows, Verilator three known rows; depth-32 native/Verilator both outcomes and default/value/partial/explicit aggregate initialization; 43 rejection cases with publication protection; generated outside-loop assertion failure.
• Python unit/quickstart suite: 578 passed, 5 skipped. Strict docs, changed-file pre-commit and whitespace checks passed.

The catalog now contains 69 examples / 25 API cases and 66 / 13 systems. These are inventory counts, not full-catalog acceptance. The three remaining historical roots (multirate_compute, nested_config_types, parameterized_types) require capabilities/policy that cannot be truthfully replaced by fixed fixtures; their exact original requirements are recorded in #272.

The extra depth-32 Icarus known-value scaling probe exceeded its 600-second resource limit and was stopped after 707 seconds; its simulation was not run. It is not counted as passing. Required main Icarus X/Z coverage and depth-32 native/Verilator outcomes remain in the semantic gate. Exact evidence and reproduction are tracked in #265. Full nightly and cross-platform native release acceptance were not run.

Candidate-bound commands, hashes, independent reviews, results and performance records remain local ignored artifacts under docs/gates/logs/array-scans-20261009/ and docs/gates/logs/array-scans-final-20261010/; no tests depend on those logs.

Independent code review: APPROVE, all 36 files bound to reviewed content; architecture: CLEAR. Final local candidate: 21255887baf2e1141f9ec3c3c85812f0ae423c50. The native producer test checks exact input-derived AST sites, ordered iteration/call frames, outside-loop origin restoration and active malformed-frame rejection.