Roadmap: make Azazel credible for large Zig build graph replacement

[godofecht]

Goal

Make Azazel a credible replacement layer for large real Zig build graphs, not just a declarative toy for small examples.

Current evidence

The huge corpus runner has real build proof for:

• libxev: builds on Zig 0.16.0
• libvaxis: builds on Zig 0.16.0
• zig-gamedev: builds on Zig 0.15.2

The remaining corpus repos are blocked by concrete non-Azazel or missing-feature issues:

• zls: supported Zig dev-toolchain window
• river: host pkg-config / wayland-scanner
• mach: custom Mach Zig toolchain
• microzig: dependency fetch/temp zip failure
• capy: transitive package format drift

Roadmap

• Resolve exact/dev/custom Zig toolchains.
• Model package dependencies from build.zig.zon well enough to diagnose/fetch/replay them.
• Model generated files and run-artifact outputs.
• Model nested workspaces and subproject build packages.
• Add executable parity commands for declared target slices.
• Produce artifact-level proof: built target exists, expected generated outputs exist, tests run when available.

Done when

• At least three corpus repos have an Azazel target slice that builds through Azazel, not only upstream zig build.
• tools/huge_corpus.py --build --parity produces machine-readable results with no false claims.
• Documentation clearly distinguishes upstream-build proof from Azazel replacement parity.

[godofecht]

Created implementation issues:

• Support exact, dev, and custom Zig toolchains in corpus manifests #17 Support exact, dev, and custom Zig toolchains in corpus manifests
• Model build.zig.zon package dependencies for replacement parity #18 Model build.zig.zon package dependencies for replacement parity
• Track generated files and run-artifact outputs #19 Track generated files and run-artifact outputs
• Represent nested workspaces and subproject target slices #20 Represent nested workspaces and subproject target slices
• Make corpus parity executable, not scaffold-only #21 Make corpus parity executable, not scaffold-only
• Add first-class system dependency and pkg-config diagnostics #22 Add first-class system dependency and pkg-config diagnostics

The immediate proof target should be one small parity: ready corpus slice before claiming broader replacement parity.