Benchmark Phase 2: full suite, aggregator, committed baseline

[suleimansh]

Child of #75. Builds on Phase 1 (#79). Full suite + automated aggregation + first tracked baseline.

Goal

Run the full task set end to end, aggregate, and commit a baseline we can diff over time (same discipline as the Rudder playground/bench baseline).

Deliverables

• aggregator: median time, median interventions, success rate per (framework, task) from the raw report.json files.
• Per-task report table: framework | median time | median interventions | success rate | n.
• A committed baseline JSON + a documented re-run command.
• Complete the task set per the Benchmark: GemStack AI orchestration vs Next.js (task speed + human interventions) #75 design; include tasks where the layer gives no edge (guard against selection bias).
• A short results write-up framed around the threats-to-validity section in Benchmark: GemStack AI orchestration vs Next.js (task speed + human interventions) #75.

Out of scope

Charts/visualization are optional and can follow.