GeezerVison

Summary

Add a reproducible Criterion workload for investigating engine/context-folding performance:

cargo bench --locked --bench engine
├── init/queries/{default,no_context}
├── engine/pr998/{default,no_context}
└── shape/pr998/context

Bundles four TypeScript/TSX before/after pairs from Whiteboard PR #998 (215,152 bytes), with provenance, hashes, and their MIT license. The benchmark reuses the actual CLI component host; the only host-source edit makes a module path explicit so it can also compile in the bench target.

docs/benchmarks.md describes baseline comparisons, timing boundaries, and rebuilding bundled plugin components after changing their Rust sources. These are native CLI/Wasmtime measurements, not browser-WASM timings. No-context also removes its query contributions; isolated shaping uses fresh clones of the same default-query trees.

Evidence

• Before: no checked-in cargo bench workload for this regression; reproducing the earlier measurement required an external fixture/harness.
• After: all five benchmarks completed in an optimized build on Rust 1.99.0. Initial Criterion point estimates: query compilation 616 ms default / 100 ms without context; four-file processing 2.266 s default / 161 ms without context; isolated context shaping 1.849 s. These are local observations, not hardware-independent thresholds.
• Passed cargo bench --locked --bench engine -- --save-baseline initial and cargo bench --locked --bench engine -- --test (all five report success).
• Passed formatting, diff whitespace checks, and all eight fixture hash/byte-count checks. Preflight rejects structural fallbacks, plugin failures, and source-text changes.

Merge Danger

Door: two-way. Blast Radius: development-only benchmark/dependency/fixture additions and one equivalent module-path annotation; no intended engine behavior change. Initial compilation includes Wasmtime and parser dependencies; full sampling takes longer than the quick --test check.

AI-assisted: OpenAI Codex.