13.4 Performance benchmark suite with regression gating #924

Labels: type:perf area:pulse-core
Milestone: Hardening - cross-cutting
Effort: Medium
Depends on: none

Description

Normalization sits on the hot path of every event, and decodedData adds spec
resolution and Zod validation to that path. No benchmark exists, so a change that
halves throughput ships unnoticed until someone runs it at volume.

Acceptance criteria

• Benchmarks cover raw event → NormalizedEvent throughput, decode-with-spec throughput, watcher fan-out at 1 / 100 / 1000 watchers, and cursor write cost per adapter
• Benchmarks run against the recorded CAP-67 fixture corpus for reproducibility
• Results are stored as JSON and compared against a committed baseline in CI
• A regression greater than 20% fails the build
• Updating the baseline requires a justification in the PR body, documented in the process
• docs/ARCHITECTURE.md records current numbers and the hardware they were measured on

Implementation notes

1. Fixture-driven, not network-driven - a benchmark that hits testnet measures
   the network.

Affected files

• packages/pulse-core/bench/ (new)
• .github/workflows/ci.yml
• docs/ARCHITECTURE.md