feat(routines): add typed declaration composition pipeline

[jscott3201]

Summary

• add a typed in-memory gate that derives declaration requirements, validates the bound-scalar branch, and checks caller-supplied source bytes
• return the existing detached bound projection only after every stage succeeds
• preserve each stage's existing error unchanged through requirements, binding, and syntax variants
• cover short-circuit ordering, atomic failure, repeatability, input immutability, document ordering, and deliberate syntax non-claims

Scope

The source-claim projection remains the declaration-side source of truth, while bound scalars remain the parallel ABI-plus-source output. This change adds no source acquisition, inventory trust policy, serialization, declaration evidence, class-kind/type/inheritance/dependency analysis, production routine rows, or external runtime integration.

Validation

• cargo fmt --manifest-path tools/routine-compiler/Cargo.toml -- --check
• focused declaration pipeline tests (6 passed)
• cargo test --locked --manifest-path tools/routine-compiler/Cargo.toml (95 passed)
• cargo clippy --locked --manifest-path tools/routine-compiler/Cargo.toml --all-targets -- -D warnings
• full Python lint suite (189 passed) and production lints
• G36 source inventory (431 release / 693 development files)
• G36 declaration, routine, verifier, and full scenario gates
• dataset harness (32 passed) and replay assertions
• generated book build

[jscott3201]

Cycle 1 consolidated review: PASS at 94a6aa069adfc01b5a73bad05e6febcb2496f409.

Both independent reviewers were blocker/major-clean, with no deferred findings. The composition preserves the required fan-out, stage order, exact underlying errors, input immutability, and atomic success boundary.

Exact-head build and verify checks are green.