fix(ci): make benchmark dispatch build and run measurements

[marcus-pousette]

What

Make the dedicated benchmark workflow reliably reach measurements and retain evidence when a benchmark fails.

• Resolve the PR number from the workflow_dispatch payload.
• Build both checked-out workspaces before running their benchmark scripts.
• Unset generic CI only for dedicated benchmark commands, so raw Playwright benchmark specs execute.
• Add a safe dispatch selector for full or focused web measurements; /bench still runs the full suite.
• Aggregate and upload partial JSON, command logs, and Playwright traces even after a base/head benchmark failure.
• Publish the available results and each command outcome, then leave the job failed when a benchmark failed.

Why

The previous workflow had three dormant gates:

1. It read an action-local input that was never supplied, so manual dispatch failed before checkout.
2. It installed dependencies without building workspace packages and generated wa-sqlite assets.
3. GitHub Actions set CI=true, while the raw browser benchmarks intentionally skip under normal CI.

Two real #164 dispatches exercised both outcome paths:

• Run 29202455068 retained 65 partial rows and logs, updated the PR comment, and stayed red after an OPFS failure.
• Run 29203415286 completed both sides, uploaded the full artifact, commented the results, and finished green.

The failed run led to the independent OPFS pathname fix in #213, which is exactly the kind of evidence the old workflow lost.

Scope / fast path

• One commit; one workflow file, +71/-6.
• No production, dependency, lockfile, benchmark, or test-source changes.
• Suite names map through a fixed allowlist; no arbitrary script name reaches the shell.
• Fork gating and same-repository security behavior are unchanged.
• Normal CI still skips nondeterministic performance measurements.
• Focused web dispatch avoids unrelated benchmark families during browser-only validation.

Verification

• Workflow YAML parses.
• Prettier and git diff --check pass.
• Shell pipe failures propagate through tee.
• Invalid suite values are rejected.
• Real dispatches verified input resolution, symmetric builds, raw browser execution, focused selection, complete and partial artifact upload, PR commenting, and final success/failure propagation.
• Normal GitHub CI is green.

[marcus-pousette]

The first workflow_dispatch reproduced the broken PR input lookup. After switching to the dispatch event payload, the second validation run resolved PR #164 successfully and reached checkout: https://github.com/cybersemics/treecrdt/actions/runs/29200644118