fix(bench): give benchmark.ts's Full build metric a warmup-then-median too

[carlos-alm]

Summary

• scripts/benchmark.ts's buildTimeMs (feeding BUILD-BENCHMARKS.md) measured a single cold buildGraph call with no warmup and no median at all — a worse instance of the exact class of bug PR fix(bench): give Full build the same warmup as the other incremental metrics #2551 already fixed for the sibling incremental-benchmark.ts's own fullBuildMs (feeding INCREMENTAL-BENCHMARKS.md).
• Full build is the first buildGraph call in this worker process's lifetime, so it's the metric most exposed to one-time NAPI/rusqlite/OS-page-cache cold-start cost — without warmup, that cost lands inside the single measured sample instead of being absorbed beforehand.
• Mirrors incremental-benchmark.ts's already-established methodology exactly: 2 warmup runs (deleting the DB before each, matching a real full build), then a median of 5 timed runs via timeMedianWithValue, which also preserves buildResult for the phases breakdown used further down in the script.
• Filed as a courtesy follow-up from Pre-publish benchmark gate fails non-deterministically on unrelated PRs #2436's investigation (alongside follow-up: benchmark gate noise is cross-invocation (CI runner variance), not within-process — needs job-level replication or dedicated runner (#2436) #2550, still open), and this issue's own text gave the exact fix shape to apply.

Closes #2549

Test plan

• scripts/ is excluded from both tsconfig.json and the biome lint scope (src/+tests/ only), and no existing unit test imports or executes scripts/benchmark.ts directly (confirmed: tests/unit/hub-selection.test.ts's reference to scripts/benchmark.ts is a synthetic fixture filename string, unrelated to the real file's content) — matching PR fix(bench): give Full build the same warmup as the other incremental metrics #2551's own precedent for this exact class of fix (no test file changes there either, since verifying requires an actual full build's timing behavior, impractical for the fast suite).
• Ran node --experimental-strip-types --import ./scripts/ts-resolve-loader.js scripts/benchmark.ts directly against this repo — completed successfully for both engines (wasm and native), producing valid buildTimeMs, perFile.buildTimeMs, and phases output for each.
• npx vitest run tests/unit/hub-selection.test.ts — 6/6 pass, confirming the unrelated fixture reference is unaffected.
• npm run lint clean.

[greptile-apps]

Greptile Summary

The PR changes the full-build benchmark from one cold sample to two warmups followed by the median of five timed runs, while retaining the build result associated with the median sample for phase reporting.

• Adds a dedicated five-run full-build sample count.
• Deletes the benchmark database before every warmup and timed full build.
• Uses timeMedianWithValue so timing and phase data come from the same selected run.

Confidence Score: 5/5

The PR appears safe to merge with no actionable correctness, security, or quality issues identified.

The timing helper preserves the result from the median-duration invocation, and the repeated database-reset methodology matches the repository’s existing sibling full-build benchmark.

Important Files Changed

Filename	Overview
scripts/benchmark.ts	Replaces the single cold full-build measurement with the established warmup-and-median methodology without introducing an actionable defect.

_{Reviews (1): Last reviewed commit: "fix(bench): give benchmark.ts's Full bui..." | Re-trigger Greptile}