t27b: run-time i128, stage 2 -- annotated locals, pair constants, @intCast, checked + - * and @divTrunc (source_text) #8763

Stage 2 of the run-time i128 in t27b (see #7915; the owner chose option A, the register pair, on 2026-10-10). Stage 1 was #8736 / #8750.

Goal

specs/tri/t27b/source_text.t27 passes t27b test --check with run-time asserts, and its ledger row is set to pass.

What it needs (measured in the #7915 comment and the stage-1 report)

• i128 accepted on an annotated local and on a reassigned var.
• The run-time module constants NO_INT, I128_MAX (2^127-1) and U64_MAX (2^64-1) as pair constants, instead of wide_plan.t27 refusing a wide constant beside a run-time value.
• @intCast from u8 to i128 (a lossless widening).
• Checked +, - and *, each trapping with "integer overflow" at the operator's line:
  • + as adds/adcs and - as subs/sbcs, the overflow from the high word's flags;
  • * as mul/umulh plus the cross terms, with an exact overflow check.
• @divTrunc through a 128-bit division routine written in t27 and compiled by t27b (never hand Rust). It traps with "division by zero" on a zero divisor, and minInt / -1 traps with "integer overflow".
• The interpreter (eval.rs, through eval_arith.t27's arith) doing exact 128-bit arithmetic without overflowing its own i128, so the JIT and the interpreter still agree.

Where the decisions live

specs/tri/t27b/i128_plan.t27 (the instruction sequences, their overflow conditions, a model that runs them, the division routine), eval_arith.t27, wide_plan.t27 and cmp_op.t27, generated with t27c gen-rust. The hand-written glue stays net <= 0 (check_budget, see #8236), paid for by porting codegen.rs functions into plans.

Acceptance

• source_text passes t27b test --check; a conformance spec under specs/tri/t27b/conformance/ passes under the reference and t27b; _expect_trap cases cover overflow on +, - and *, a zero divisor and minInt / -1, interpreter and JIT agreeing on each site.
• Corpus before/after: only intended files move; jit_interp_mismatch stays 0.
• check_budget and check_all exit 0.

Stage 3 (eval_arith: parameters in pairs, struct fields, @rem, the bitwise operators, shifts) is not part of this issue.