docs(memory): define durable learning boundary

[627150795]

Problem

Repository instructions, Session transcripts, compaction summaries, Skills, caches, and durable memory have different owners and lifecycles. Treating them as one memory store can persist stale or untrusted text.

Value

Define a manual-first, project-scoped durable-learning contract with explicit record classes, provenance, authorization, expiry, recall, and Skill-promotion boundaries.

Approach

• Add docs/design/PROJECT_DURABLE_LEARNING_CONTRACT.md.
• Separate facts, preferences, experiences, Skill candidates, Session history, indexes, and repository instructions.
• Require explicit authorization, project/worktree identity, bounded recall, independent receipts, and auditable stale or disputed states.
• Keep automatic writes, automatic prompt injection, vector retrieval, and Skill mutation out of scope.

Validation

• git diff --check
• bun run lint
• bun run typecheck

Impact

Documentation only. This focused manual-first design slice for #167 does not add a memory database or automatic recall.

Addresses #167

[tt-a1i]

Thank you for writing down this candidate. We are closing it under the future-capability disposition in #277. The PR remains the preserved research source, but OpenPI will not promote a protocol-only or candidate design into a repository contract before there is an active implementation or experiment with a fixed owner, revision, verifier, and acceptance evidence. This is not a content rejection, and no rebase is requested. Restart the work from the linked source Issue when that evidence frontier is ready.