Document lessons from the Hyperkernel experiment

[hnordt]

Summary

• add a historical retrospective of the external hnordt/hyperkernel-experiment repository, pinned to the evaluated revision
• record the kernel, benchmark, event-log, and time-travel UI lessons that should inform the official implementation
• distinguish reusable ideas from experimental contracts and implementation choices that must not be promoted
• index the retrospective in docs/spikes/README.md and record the experiment's intended archival disposition

Why

The experiment made the proposed architecture tangible and produced useful evidence about atomic event and projection writes, constrained SQL, retained-history performance, and read-only time travel. Capturing that evidence before the experimental repository is archived prevents it from becoming an undocumented second implementation or an accidental supported contract.

Impact

This is documentation-only. It changes no runtime behavior, public SDK, database schema, or supported Kernel, Extension, or Experience contract.

Change classification

• Experience
• The retrospective discusses evidence across Kernel, Extension, and Experience boundaries, but explicitly does not define or modify their supported contracts.

AI usage disclosure

• Tool: OpenAI Codex with multiple collaborating agents
• Parts using AI: planning and documentation authoring; existing tests and benchmarks were executed by agents; no test code was authored
• Overall usage level: automated
• Human review remains required. Agent review and automated checks do not constitute approval.

Verification

Official repository:

• npm run check
• npm run lint
• npm run test — 1 unit test and 1 Playwright test passed
• npm run build
• git diff --check

Experiment repository at 48527996a23c5dfde85426d7350bc7928f78a616:

• deno task check — 32 tests passed
• repeated 9-sample lifecycle benchmarks at 1,000 and 10,000 Todos
• deno task todo:test — 17 tests passed

Uncertainty and risks

• Benchmark results are environment- and workload-specific; the new measurements are retrospective evidence, not a performance guarantee.
• The current-head benchmark results were produced ad hoc and are not committed to the archived experiment.
• The first sandboxed npm run test attempt was blocked from opening a local port with EPERM; the complete suite passed when rerun with local-port permission.
• This PR introduces no security, compatibility, recovery, or migration behavior. The document identifies unresolved risks in those areas for future design work.
• Archiving the experiment repository is an external follow-up; this PR records the intended disposition but does not archive it.