chore(ci): enhance ci performance profiling and add new scripts

.agents/skills/profile-ci/SKILL.md

@@ -0,0 +1,138 @@
+---
+name: profile-ci
+description: Run and analyze Elements cold CI performance profiles and propose evidence-backed build, test, lint, and dependency-graph improvements. Use whenever asked to profile or benchmark pnpm run ci, rerun ci:profile, create a CI performance audit, compare CI timings, find bottlenecks or the completion path, investigate a CI performance regression, or recommend measured CI/build optimizations.
+---
+
+# Profile CI
+
+Produce a repeatable cold-CI profile, explain what controls wall-clock completion, and turn the evidence into prioritized, testable recommendations.
+
+## Required context
+
+1. Read the repository `AGENTS.md`.
+2. Read `projects/internals/BUILD.md`.
+3. Inspect the current root `ci:profile` script and `projects/internals/ci/ci-profile.js`. Treat the script as the profiling source of truth.
+4. Read a project's `DEVELOPMENT.md` before running project-specific commands when that file exists.
+
+Do not copy profiling logic into this skill. Update the repository profiler when its behavior needs to change.
+
+## Choose the workflow
+
+- Run a new profile when the user asks to rerun, benchmark current changes, refresh an audit, or verify an optimization.
+- Analyze existing `.metrics/ci-profile.{json,md}` artifacts without rerunning when the user asks only for interpretation and the artifacts match the intended commit and worktree state.
+- Create recommendations without implementing them unless the user also asks for the changes.
+
+## Profile safely
+
+Run all repository commands through mise.
+
+1. Inspect `git status --short`.
+2. Preview ignored files that reset would delete with `git clean -ndX`.
+3. Stop and ask before profiling if that preview includes user data, local assets, secrets, or other non-reproducible files. The profiler runs `pnpm run ci:reset`, which deletes ignored files and reinstalls dependencies before every sample.
+4. Never stash, commit, discard, or clean tracked changes merely to make the profiler accept the worktree.
+5. Use the clean command when the worktree is clean:
+
+   ```shell
+   mise exec -- pnpm run ci:profile
+   ```
+
+6. When the dirty changes are the intentional subject of the audit, preserve them and run:
+
+   ```shell
+   CI_PROFILE_ALLOW_DIRTY=1 mise exec -- pnpm run ci:profile
+   ```
+
+   Record the dirty-worktree condition in the report. Do not use this override for unrelated or unexplained changes.
+
+Allow all three cold samples to finish. Dependency installation or browser setup can require network access. If a run fails, inspect the copied `.metrics/ci-profile-run-*.log` and reset logs, report the incomplete profile, and do not invent missing samples.
+
+## Verify the artifacts
+
+The profiler writes ignored artifacts under `.metrics/`:
+
+- `ci-profile.json`: structured metadata, run durations, and per-script samples
+- `ci-profile.md`: generated method and top-ten summary
+- `ci-profile-run-{1,2,3}.log`: complete CI logs
+- `ci-profile-reset-{1,2,3}.log`: reset and install logs
+
+Before analysis, confirm:
+
+- all three runs exist;
+- every run has zero incomplete scripts;
+- each ranked script has three samples;
+- commit, dirty state, tool versions, CPU, and memory describe the intended environment;
+- the generated Markdown agrees with the JSON.
+
+Use JSON as the numeric source of truth. Keep full precision during calculations and round only for presentation.
+
+## Analyze wall-clock relevance
+
+The slowest command is not automatically the critical path.
+
+1. Rank leaf scripts by median duration from `ci-profile.json`.
+2. Inspect the end of each run log to identify the scripts that consistently complete last.
+3. Trace those scripts through their Wireit `dependencies` in the relevant `package.json` files.
+4. Separate:
+   - final or near-final dependency branches;
+   - long parallel work that consumes CPU, memory, filesystem, or browser capacity;
+   - upstream work that delays a final branch;
+   - composite commands whose phases need separate timing.
+5. Inspect each candidate's command, configuration, file count, output, output consumers, and serialization settings before suggesting a change.
+6. Compare with the existing audit only when its environment and method are compatible. Describe before/after changes as concurrent cold-CI measurements, not isolated causal proof.
+
+Read [the optimization playbook](references/optimization-playbook.md) when generating recommendations or designing follow-up experiments.
+
+## Create the audit
+
+Create or refresh `projects/internals/ci/CI-PERFORMANCE.md`. Use `.metrics/ci-profile.md` as generated evidence, not as the finished audit.
+
+Include:
+
+1. Frontmatter and generation context
+2. Executive summary with median CI time and comparison when available
+3. Method, environment, and all run results
+4. Current top-ten scripts
+5. Completion-path analysis
+6. Numbered open findings
+7. Verified completed changes when a prior audit exists
+8. Recommended experiment order
+9. Raw artifact location
+
+For every finding, provide:
+
+- **Evidence:** measurements plus current configuration or dependency facts
+- **Recommendation:** one bounded change or experiment
+- **Validation:** output-equivalence checks, metrics to compare, and resource checks
+- **Confidence:** confidence in the diagnosis and in the proposed approach
+
+Keep completed work out of the open priority list. Mark an item verified only when the configuration changed and a comparable full profile confirms the result.
+
+## Recommendation guardrails
+
+- Do not add CI-share percentages together; scripts overlap.
+- Do not claim a critical path from rankings alone.
+- Do not recommend more concurrency without checking isolation, shared state, ports, output paths, report merging, memory, and machine-wide contention.
+- Do not remove a quality check unless the same coverage remains in another required workflow.
+- Prefer one-variable experiments and compare at least three samples when variance matters.
+- Distinguish targeted project benchmarks from full concurrent CI results.
+- Treat cache optimization separately from this cold profile because the profiler disables Wireit caching.
+- Label estimates and inferred causes. Do not present them as measurements.
+- Keep dependency upgrades separate from performance changes unless the upgrade is the explicit experiment.
+
+## Check the report
+
+Run:
+
+```shell
+mise exec -- pnpm exec prettier --write projects/internals/ci/CI-PERFORMANCE.md
+mise exec -- pnpm exec vale --config .vale.ini projects/internals/ci/CI-PERFORMANCE.md
+git diff --check
+```
+
+If `projects/internals/ci/ci-profile.test.js` exists, also run:
+
+```shell
+mise exec -- node --test projects/internals/ci/ci-profile.test.js
+```
+
+Do not rerun the full CI merely to validate the report: the profiler already completed it three times. Report the median, comparison, top remaining opportunities, artifact paths, and validation results to the user.

.agents/skills/profile-ci/references/optimization-playbook.md

@@ -0,0 +1,108 @@
+# CI optimization playbook
+
+Use this reference after collecting a valid profile. Select approaches from current evidence; do not list every approach in every audit.
+
+## Selection order
+
+### Remove unconsumed or duplicate work
+
+Look for bundle visualizers, duplicate test suites, unused report formats, repeated downloads, redundant generation, and work repeated in both a main job and a dedicated job.
+
+Verify all output consumers before removal. Preserve required test, release, documentation, and deployment coverage.
+
+### Improve the completion branch
+
+Trace the scripts that finish last through Wireit dependencies. Consider:
+
+- starting independent prerequisites earlier;
+- removing dependencies that do not represent real input requirements;
+- narrowing broad dependency fan-in;
+- reducing work in the final command;
+- avoiding a second copy, transform, or package pass.
+
+Do not decouple a dependency until output and runtime behavior prove that the downstream task does not need it.
+
+### Split composite commands
+
+Split sequential phases when one command hides attribution or forces broad cache invalidation. Give each phase accurate Wireit files, outputs, dependencies, and environment.
+
+Splitting improves observability and caching. It does not reduce cold wall time by itself unless the graph can safely overlap phases.
+
+### Add controlled parallelism
+
+Look for one-worker test suites, disabled file parallelism, sequential linting, and independent generated targets.
+
+Prefer process-level shards when tools share browser state or globals. Give every shard separate ports, browser profiles, coverage directories, screenshots, JUnit/JSON output, and temporary files. Merge reports deterministically and verify identical totals and thresholds.
+
+Measure peak memory and full CI time. A faster package command can slow the pipeline by starving sibling work.
+
+### Reduce transformed or generated inputs
+
+Use module counts, page counts, plugin timings, and repeated transforms to find large input surfaces. Consider:
+
+- deduplicating identical generated modules;
+- externalizing safe shared assets;
+- omitting development-only modules from production;
+- avoiding repeated full-document scans;
+- narrowing entry points and globs;
+- generating shared metadata once.
+
+Require output-equivalence checks for routes, bundles, screenshots, metadata, and runtime behavior.
+
+### Improve cache boundaries
+
+Audit Wireit `files`, `output`, `dependencies`, `cascade`, environment variables, and package-lock inputs. Split tasks whose unrelated inputs invalidate expensive work.
+
+Measure cache-hit behavior separately. The cold profiler sets `WIREIT_CACHE=none`, so cache changes cannot explain its results.
+
+### Use tool-specific diagnostics
+
+Inspect the installed tool version and local configuration before choosing flags.
+
+- Wireit: use command start/finish order and dependency declarations.
+- Vite or Rolldown: inspect transformed module counts and plugin timing warnings.
+- Eleventy: compare written files, templates, transforms, and Vite phases.
+- Vitest: inspect file/test totals, workers, concurrency, isolation, retries, browser providers, and reporters.
+- ESLint: inspect rule timing, file counts, cache state, and supported concurrency modes.
+- Coverage: identify report consumers and the cost of collection, transformation, serialization, and report generation.
+
+## Experiment record
+
+For each experiment, capture:
+
+- hypothesis;
+- exact configuration difference;
+- baseline and candidate commands;
+- sample count and environment;
+- median, minimum, maximum, and variance;
+- output-equivalence checks;
+- test and coverage totals;
+- peak memory or other resource constraints;
+- targeted timing and full-CI timing;
+- decision and confidence.
+
+Change one performance variable at a time where practical.
+
+## Confidence rubric
+
+- **High diagnosis confidence:** repeated measurements and configuration directly explain the serialization or work.
+- **Medium diagnosis confidence:** evidence is consistent, but concurrent contention or hidden tool behavior remains.
+- **Low diagnosis confidence:** the signal is noisy, appears only once, or lacks configuration support.
+- **High approach confidence:** a low-risk removal or configuration change preserves verified outputs.
+- **Medium approach confidence:** the approach needs isolation, report merging, or dependency validation.
+- **Low approach confidence:** the expected gain is speculative or depends on undocumented behavior.
+
+State diagnosis confidence separately from approach confidence when they differ.
+
+## Reject weak recommendations
+
+Do not recommend:
+
+- deleting caches to make a warm build faster;
+- adding all script durations to estimate total CI time;
+- increasing every worker count to the CPU count;
+- removing tests solely because they are slow;
+- moving a check to another workflow without confirming that required status checks include it;
+- claiming an isolated benchmark equals full CI savings;
+- upgrading dependencies and changing orchestration in the same performance experiment;
+- optimizing a top-ten parallel script while ignoring the branch that completes last.

config/vale/styles/config/vocabularies/Elements/accept.txt

@@ -9,6 +9,7 @@ ESLint
 Wireit
 pnpm
 Vite
+Rolldown
 Vitest
 WCAG
 ARIA
@@ -30,8 +31,10 @@ Chromium
 Playwright
 OWASP
 CDN
+CPUs
 polyfill
 polyfills
+profiler
 entrypoint
 entrypoints
 postcss
@@ -72,15 +75,19 @@ popover
 popovers
 shortcode
 shortcodes
+passthrough
 server-driven
 single-column
 multi-column
 Heatmap
 heatmap
 heatmaps
+hotspot
 datalist
 debounce
 debounced
+deduplicating
+deduplication
 dropdown
 dropdowns
 checkbox
@@ -124,8 +131,10 @@ dev
 async
 Rollup
 Webpack
+Turbopack
 subheader
 tsconfig
+typecheck
 tsx
 TSX
 jsx
@@ -147,6 +156,7 @@ svg
 SVG
 scrollable
 parallelization
+sharding
 minimizable
 datagrid
 ESBuild
@@ -174,6 +184,7 @@ inlined
 Inlined
 globals
 gitignored
+untracked
 exportparts
 downleveling
 downlevel

package.json

@@ -20,6 +20,7 @@
     "ci:all": "PAGES_BASE_URL=/elements/ pnpm run ci && PAGES_BASE_URL=/elements/ node ./projects/internals/ci/cache-validate.js ci",
     "ci:reset": "pnpm run clean && pnpm i --frozen-lockfile --prefer-offline",
     "ci:validate": "wireit",
+    "ci:profile": "node ./projects/internals/ci/ci-profile.js",
     "lighthouse": "wireit",
     "clean": "git clean -dfX",
     "format": "wireit",
@@ -102,6 +103,7 @@
         "./projects/starters/vue:ci",
         "./projects/styles:ci",
         "./projects/themes:ci",
+        "./projects/internals/ci:ci",
         "./projects/internals/design:ci",
         "./projects/internals/metadata:ci",
         "./projects/internals/patterns:ci",