Bool gate costs ~46 Ir per builtin call (+2.9% on a builtin-heavy loop): make it structural #1654

Measured in #1647 round 6 (callgrind, deterministic). The boolean type's per-call bool gate (eigs_call_builtin scanning arguments for a bool in a position the builtin does not declare) costs about 46 Ir per builtin call: +2.9% on a 50k-iteration loop of 4 builtin calls (368.1M → 378.8M Ir). On/off attribution: with the inline scan disabled, the loop measures 368.1M, so the scan is the entire cost. It is within the merge gate (5%), and the basic loop benches sit at +0.33%. But builtin-heavy consumers (DMG, iLambdaAi, eigen-sheet) pay it on every call.

Options (from the round-6 builder)

• A bool-content flag on lists, set when a bool is stored, so the scan becomes one flag test.
• Do the check when the argument list is built from slots, where the tag is already in hand.
• A per-builtin declared-no-bool bit that skips the scan entirely for builtins that take no bool anywhere.

Done when

• The builtin-heavy loop's overhead against the pre-Decision: a boolean type (#1286 booleans row → CHANGE) #1637 baseline is ≤ 1% (paired callgrind), with the Decision: a boolean type (#1286 booleans row → CHANGE) #1637 fuzz (tests/test_bool_fuzz.sh) still at 0 violations and num_read_check.sh passing.
• DMG's CodSpeed instruction count shows no regression against the pre-Decision: a boolean type (#1286 booleans row → CHANGE) #1637 pin.