[backend] Productise code generation

Architecture

This is a fresh implementation from origin/main; the earlier NbE and SSA experiments are scoping evidence only.

flowchart LR
    checked[Checked tree] --> nbe[NbE tree]
    nbe --> nbePasses[NbE passes]
    nbePasses --> javascriptTree[JavaScript tree]
    nbePasses --> ssa[SSA tree]
    ssa --> ssaPasses[SSA passes]
    ssaPasses --> javascriptTree
    javascriptTree --> javascript[JavaScript]

The shared JavaScript crate owns a backend-neutral JavaScript tree and renderer. It does not depend on the NbE or SSA IR crates.

Scope

Add CLI/build integration, output layout, linker/module discovery, and source mapping after both paths execute the corpus. Crate-level build queries are prerequisites and are not deferred to this issue. Decide whether both paths remain user-facing or the direct path becomes an oracle/reference backend.

Acceptance criteria

• A clean checkout can compile and execute a multi-module fixture without test-only host state.
• Output and foreign-module discovery have a documented layout.
• Source maps cover representative generated forms.
• All backend categories pass unfiltered with no pending snapshots.

Dependencies

• #338
• #341

Start by reading dependencies #338 and #341, then trace the checked-tree to JavaScript flow and the shared JavaScript crate described in the architecture. Define the CLI/build, output layout, linker and module-discovery boundaries before implementation. Done means a clean checkout runs the multi-module fixture, documents output and foreign-module discovery, produces representative source maps, and passes all backend categories without pending snapshots.