bench: record Phase 0 run for task-001-tags (both apps PASS)

[suleimansh]

First end-to-end Phase 0 run of the "our AI vs Next.js" benchmark (#75, #78).

What this is

The actual measurement run on top of the merged Phase 0 harness. Not new product code, just recorded results plus the agents' solution patches. The baseline apps stay pristine (the Phase 0 starting point); only benchmarks/runs/ is added.

Method

• One autonomous AI coding agent per app, same agent type and model both sides (fair-harness rule), same task spec verbatim, no human in the loop.
• Harness validated first: both committed baselines fail the gate identically (same 5 checks) -> a real symmetric gap.
• Independently re-verified by hand against fresh DBs + restarted servers after the agents reported PASS.

Results

App	Status	Time	Interventions
bench-app-next (vanilla Next.js)	PASS	139s	0
bench-app-gemstack (Vike + @gemstack/ai-sdk)	PASS	145s	0

All 14 contract checks pass on both. Diffs comparable (~130-150 lines each).

Reading

The rubric runs end to end and the gate is objective and reproducible, so Phase 0's goal is met. A plain CRUD-extension task does not exercise the orchestration layer, so it does not differentiate here (both fast, zero interventions). The intervention metric will only show signal on AI-integration / multi-step / refactor tasks. Phase 1 (#79) should weight the task set toward those.

See benchmarks/runs/2026-06-28-task-001-tags/results.md for the full log and reproduce steps.

[suleimansh]

Not merging. The benchmark is relocating to its real home, suleimansh/vike-data, where the extension family it now measures (vike-auth, vike-data, vike-notifications, vike-stripe) actually lives. This Phase 0 run proved the method (harness + rubric work; a plain CRUD task doesn't differentiate) — that learning carries over to the vike-data benchmark. Closing here rather than landing an AI-only run log on a repo the benchmark is leaving.