Imported generics and return validation

The canonical contracts are in language.md, sections Size-Generic Functions, Constants, and Return. This change fixes executable source semantics; it does not extend the nox match implementation or imply that unsupported proof relations are complete.

Reproduced defects and corrections

1. Public generic signatures were exported as ordinary signatures with unknown dimensions evaluated as zero. Imported explicit calls failed with “not generic” and shape mismatches. ModuleExports.generic_functions now carries unresolved size parameters/types separately, including the canonical owner identity and lexical signature context. Ordinary exported functions no longer contain fabricated zero-sized generic signatures.
2. Generic bodies were skipped without subsequent concrete body checking. Project preparation now resolves exact call sites, allocates deterministic collision-free concrete function names, and materializes each function in its defining module. It repeats type checking before lowering. Nested calls infer sizes from their actual concrete argument shapes; identical source spans in different concrete caller functions remain separate. Private helpers, cfg selection, nominal types, and lexical constants stay with the owner. Both nox and Trisha consume ordinary checked concrete functions.
3. Ordinary function explicit returns and terminal values were never checked against the declared return type. They now are, including missing fallthrough results, terminal if/match branch values, and fixed nonempty versus empty loops. Parameter/local shadowing cannot turn a runtime condition into a module-constant termination claim. A shared AST terminal-return normalization preserves expression spans and scopes for checking, formal analysis, and nox lowering. The nox lowering correction and its proof regressions are recorded by the parent agent separately.
4. Formal scalar analysis discarded terminal if branch values and could model their result as zero. It now consumes the same return normalization. A real Z3 regression rejects ensures(result == 0) for a terminal 7/9 conditional.
5. Bare module constants in shared TIR fell through to a stack duplicate. They now emit their constant value after local/spilled bindings take precedence. Constant references preserve declared Field/U32 types; U32 initializers are range checked. Generic expression substitution preserves nominal module constants while substituting size parameters. Constants are registered before signatures so forward declarations are valid.

Dimension substitution rejects unresolved names and arithmetic overflow. Literal out-of-bounds array reads, wrong generic arity/shapes, duplicate size parameters, and recursive instantiation are rejected. Expansion has explicit limits of 1,024 instances and 128 rounds. Recursion detection respects cfg. Single-source compile, TIR, check and editor/project checking use the same concrete preparation rather than bypassing body checking.

Validation

• tests/imported_generics.rs: five tests cover actual nox execution in both profiles, nested imported entry shapes, lexical dimensions, malformed calls, invalid generic bodies, ordinary return errors, missing paths, nominal U32 constants/range, and single-source API parity.
• Independent Trisha rs/tests/imported_generics.rs: five tests pass in both profiles. Cases include wide 20-word results, nested inferred sizes, two owner modules with the same function name, private helpers, cfg, module constants shadowed by size parameters and caller variables, and user names colliding with proposed specialization labels. Owner logs: /tmp/trisha-imported-generics-review.log.
• The owner also ran a genuine default-security Triton proof for input [3,5] and output 42, rejecting changed input, output, and program digest: /tmp/trisha-imported-generics-proof.log.
• The 12 real-Z3/formal/native tests pass with the terminal-if regression: /tmp/generic-return-targeted.log.
• The first complete workspace neural run passed 882 tests, zero failures, zero ignored tests and zero Rust warnings across 19 test reports: /tmp/generic-return-workspace-neural-final.log. The final repeat after forward constants and canonical Field-condition handling passed 883 tests, zero failures/ignored/Rust warnings, across 19 reports: /tmp/generic-return-workspace-neural-final3.log.

The full library check initially exposed 19 module failures caused by the old checker treating terminal branch values as unit. Once terminal values were modeled correctly, only two source files required actual repairs: std/compiler/codegen.tri and std/compiler/typecheck.tri had unit helpers leaving Field-returning calls as their terminal values. The typed owner made those discards explicit; the checker was not weakened to accept them.

Historical release archives and proof receipts are not represented as rebuilt by these tests. Current release artifacts and the full proof corpus remain the parent agent's coordinated release checks.

Final module check: 48/48 core-library and production SDK modules passed trident check --target neptune using current API-3 owner discovery. The current inventory is /tmp/generic-return-inventory/results.json; its historical JSON key audit_neptune records this run's check exit code, not a formal proof.

Canonical Field handling received an additional boundary regression: scalar program literals are reduced modulo Goldilocks for formal evaluation and branch termination inference; compile-time sizes are not reduced. Symbolic negation of zero now yields canonical zero. Trisha's textual immediate normalization is an owner correction tested separately; the earlier parser rejection must not be counted as successful Field-literal execution.

Coordinated frozen-source repeat: cargo test --release --workspace --features neural --locked exited 0 after the final shared-TIR intermediate-value fix and nox canonical Field regression. 884 tests passed, zero failed/ignored, zero Rust warnings, 19 test reports, including the 49-test nox surface suite. Log: /tmp/trident-final4-workspace-reviewed.log. Only this report was edited after the run; no implementation changes were made by this agent.