[Docs] OKF (Open Knowledge Format) study; improvements #96-#98

[codenamev]

Summary

Adds a /study-repo analysis of Google Cloud's Open Knowledge Format (spec v0.1 draft) — a vendor-neutral spec for knowledge as markdown files with YAML frontmatter in directory bundles.

Headline finding: OKF and ClaudeMemory are complementary, not competing. OKF specifies an interchange surface with no engine (no truth maintenance, no queryable provenance, no ranked retrieval); ClaudeMemory is an engine whose only interchange surface is one generated monolith. OKF is what our publish layer could speak — and must never be our store of record.

Changes

• docs/influence/okf.md — full analysis: spec structure, permissive consumption model, bottom-up index regeneration, augmentation guard (loss-refusing writes), tool-enforced crawl budgets, four-gate reference mint test, comparison table vs ClaudeMemory, prioritized adoption opportunities.
• docs/improvements.md — new "OKF Study (2026-07-16)" section:
  • #96 ⭐ claude-memory export --format okf — per-fact markdown bundle with provenance-derived citations, generated index.md progressive disclosure, log.md from memory.changes. The lock-in answer.
  • #97 ⭐ Corrective re-prompt errors + loss-refusing guards in memory.store_extraction — error shape: invariant → what to preserve → which tool to re-call. Targets Layer-2 extraction loss (#72). Loss guard starts warn-mode.
  • #98 ⭐ Four-gate reference mint test in the SessionStart distillation prompt — targets the documented reference-junk / distiller-hallucination pattern.
  • Medium items (log.md changelog in the snapshot, sorted-pair naming for symmetric relationships) and Features to Avoid (per-call LLM synthesis, files as store of record, global tool state, CDN-dependent artifacts).

Items were renumbered from #76–#78 to #96–#98 during rebase: the 2026-06-30 and 2026-07-09 re-study cycles on main had already consumed #76–#95.

Docs only — no runtime code changes.