Independent terminal-return review

Date: 2026-09-12. Status: reviewed cases pass; owned changes stable. No heavy proof was executed for this gate. Frozen archives were not changed.

Scope: ordinary and imported Trisha functions ending in If/Match, nested locals, explicit early returns, aggregate results wider than the register window, and intermediate branch/loop value disposal. Independent tests live in rs/tests/terminal_returns.rs; the existing early_return.rs covers additional explicit exits. Root separately reviews native Nox terminal normalization.

Finding and correction

The broad Trisha workspace run reported four compile failures in compiler_pipeline and triton_audit_stdlib after ERROR markers correctly became fatal. Shared TIR build_stmt compared result widths of every IfElse, including intermediate statement branches. Such branches can legitimately have distinct tail value widths: those values are evaluated for effects and discarded.

The owner authorized the reviewer to correct this boundary. The change in trident/src/ir/tir/builder/early_return.rs adds a recursive value-context pass; functions.rs applies it to a cloned function body before existing return lowering. Only a final If/Match in a value context forwards branch tails. Intermediate If/Match and all loop bodies turn their tails into ordinary Expr statements, which evaluate and discard the result. Explicit Return nodes remain untouched for the existing return-slot/flag mechanism. Function tail expressions retain their return role. No ERROR guard or actual terminal width check was weakened, and the pass does not add return-slot overhead to every function.

Actual execution coverage

Three persistent tests execute the real Triton VM with an independent 100,000 instruction cap in both debug and release source profiles:

1. Nested terminal If, local shadowing, explicit early return and caller sentinel: inputs0/1/7 return11/19/26 followed by unchanged97.
2. Imported terminal Match with nested terminal If and explicit return: four branches return a22-word aggregate, preserve every distinct array element and the caller's sentinel. Pair results are[11,13], [17,19], [23,29], [43,31].
3. Intermediate If and Match with differing tuple/scalar tails, plus loop tails: effects occur exactly once in the selected branch, loop effects occur twice, discarded values do not corrupt the final31 or caller97. Exact outputs are [11,17,23,23,31,97] and[13,19,23,23,31,97].

Verification receipt

/tmp/trisha-terminal-struct-review.log records the release test invocation:

• compiler_pipeline:3 passed (both originally failing cases repaired).
• triton_audit_stdlib:13 passed (both originally failing cases repaired).
• terminal_returns:3 passed.
• struct_literal_layout:4 passed.
• imported_generics:5 passed; its separately proven gate intentionally ignored in this non-proof invocation.

Total:28 passed,0 failed,1 intentionally ignored proof test, no compiler warnings. This is a bounded owner review, not a replacement for the root's full workspace run or independent execution-proof relation verification.

The existing rs/tests/early_return.rs was then independently rerun after the value-context change:9 passed,0 failed (/tmp/trisha-terminal-existing-returns.log). This includes nested aggregate exits, loop returns, caller preservation and skipped effects. Thus the two invocations total37 passes with the same single intentionally ignored proof gate; no proof workload was started.