[ryantm]

Summary

Complete the worker-daemon build path for consumers of dynamic derivations.
The evaluator already supports builtins.outputOf, but the worker client discarded
realized output paths and later queried the unresolved derivation for them.

• Retain and use output realizations returned by BuildPathsWithResults.
• Preserve nested dynamicOutputs in derivation JSON.
• Reject dynamic inputs explicitly on the static-only gRPC build path.
• Add a real-daemon regression covering fresh and cached builds with IFD disabled.
• Add reproducible drvgo/UPM evaluation and application recompilation benchmarks.

Direct dynamic output strings as installables remain unsupported; consumers such
as drvgo's wrapper derivations are supported. No evaluator-time builds are added.

Validation

• Fresh and cached dynamic-derivation daemon regression passed.
• 17 dynamic-related unit tests and worker result parsing test passed.
• cargo clippy --all-targets -- -D warnings, formatting, and ast-grep passed.
• Five alternating UPM runs: evaluation 0.588s versus 1.373s (2.34x); cached builds 0.603s versus 1.407s (2.33x).
• Fresh-source UPM builds and smoke tests passed, but the initial compilation samples used different job/core budgets and are not comparable. A matched-budget run failed on Determinate Nix with signal 9 in a generated forwarding derivation; no cgroup OOM was recorded. No compilation speedup is claimed.
• Reproduction scripts, raw completed samples and limitations are in bench/results/dynamic-upm.md. The VM could not enable Nix sandboxing.

Text and test evidence are used because this is a CLI/daemon protocol change.

~ written by ⠕ Replit