v2.2.0 — Pre-paper gap closure

Pre-paper gap closure. Five new ADRs (0013–0017) carry the related-work
positioning against the agent-memory literature, the failure-mode theory
of the bidirectional loop, the self-reingestion guards, the
text-observability requirement for Measure, and the harness alignment /
harness drift vocabulary. README gains a Limitations section and a
rewritten Related Work with full citations; the multi-language README
mirrors are retired in favor of English + Japanese.

Added

• Four new ADRs (0013–0016) closing the 2026-06-06 pre-paper gap analysis. ADR count 9 → 13 (gaps 0001/0006/0007 unchanged). A fifth candidate — demotion / rule retirement as Promote's structural inverse — was drafted and withdrawn the same day as premature: rules are already retired in practice through interactive curation (the gap analysis mistook a symmetry argument for observed pain), and the rule-layer-curation observation is deferred until it has a pain record:
  • ADR-0013 — positions AKC against the agent-memory / skill-learning literature (Voyager, AWM, ReMe, LangMem, Generative Agents, MemGPT, CoALA, the Externalization survey) via concede-then-locate: the operations are not novel in isolation; the human-owned loop, bidirectional growth target, and attention-side scarcity framing are the delta.
  • ADR-0014 (experimental) — names Theme 3's failure twin: gate complacency, deskilling, delegation-feedback divergence, with AKC's structural defenses.
  • ADR-0015 (experimental) — the cycle feeds on its own output: agent-written rules and notes re-enter as distillation input, so the loop can degrade (echo, grounding loss) as well as correct. General principle distilled from a Contemplative Agent production observation; the observed record is the only ground, the self-generated share of promoted artifacts stays visible (Measure), the gate is the circuit-breaker.
  • ADR-0016 — requires Measure instruments to promote agent text (reasoning, verdicts, plans) to observable events; tool-call-only instruments systematically under-report judgment-phase compliance.
• ADR-0005 addendum — "Why the gate is the contribution, not a guardrail": contrasts AKC's named-human-sign-off promotion against the autonomous-promotion default of the self-evolving-agent corpus, and frames Promote as deliberately extrinsic metacognition.
• README Limitations section and rewritten Related Work / References carrying the agent-memory prior art with full citations; mirrored in README.ja.md. Design Principle #2 sharpened with the text-observability requirement.
• akc-cycle.md Measure step sharpened: compliance is checked in agent text (stated reasoning, verdicts) as well as tool calls — the install file's only change; explanatory material stays in the ADRs.
• graph.jsonld gains nodes for the four new ADRs, loop failure modes, self-reingestion, and prior-art research lines with contrast edges. glossary gains the corresponding mechanism terms.
• ADR-0017 — harness alignment / harness drift vocabulary (ADR count 13 → 14). Establishes both terms by derivation from established literature rather than fresh coinage, after a four-domain search confirmed no single established term covers the concept. Harness alignment extends intent alignment (Christiano 2018 — citation also newly added to the README, repairing a missing attribution) to the configuration layer and across time (cf. Lehman 1980, Law I); harness drift names the failure mode in lineage with architectural drift (Perry & Wolf 1992), practical drift (Snook 2000), and agent drift (arXiv:2601.04170); the contrast is autonomous, score-driven harness optimization (Meta-Harness, arXiv:2603.28052). Quotations verified against primary texts (Snook secondary-verified). Synced across README en/ja (new "Software-evolution and alignment literature" References table + harness-engineering section paragraph), glossary, llms.txt, llms-full.txt (new Q&A; Meta-Harness citation corrected to its actual title "End-to-End Optimization of Model Harnesses"), and graph.jsonld (2 Concept nodes, 1 ADR node, 6 prior-art nodes, contrast edges).

Fixed

• Promote trigger restored to holistic judgment. The "three or more places" numeric threshold in the Promote phase was a transcription error from the original rules-ization (the first draft of docs/rules.md, 2026-03-29): the bound skills (rules-distill; learn-eval on the harness side) have always used holistic quality gates with no numeric threshold, and the section's own Trigger line was already holistic. The number also demonstrated its cost in this very repository — quoted as "recur 3+ times" in both README phase tables, silently mutating "places" into "times". Corrected across akc-cycle.md, README en/ja, graph.jsonld, and llms-full.txt: the trigger is that the repetition itself has become the signal; the Human Approval Gate (ADR-0005) remains the actual filter.
• Design-principle count unified at nine across all surfaces (ADR-0012 prose corrected; ADR-0011 carries the canonical "Genre neutrality is Design Principle #9" declaration; historical "eight" references in ADR-0009's v2.0.0 addendum annotated, not rewritten).
• Four-pattern definitions (guard / filter / judge / orchestrator) re-synced to ADR-0008 canonical in graph.jsonld, CODEMAPS, and llms-full.txt — guard validates LLM output post-generation, orchestrator means code owns the loop. CODEMAPS Invariants now pins the canonical source.
• AAP sibling references updated from "eight ADRs / five additional" to "ten ADRs / seven additional" (AAP v0.2.0 added ADR-0009/0010); concept DOI references unchanged.
• ADR-0008's dangling reference to extracted ADR-0007 removed; structural-guard vs populated-threat-model demarcation added (CLAUDE.md mechanism-only rule clarified accordingly).
• CODEMAPS file-count snapshot updated for the retired README mirrors (en + ja = 2, not 6); ADR routing extended to 0013–0016.

Removed

• Multi-language README mirrors (es, pt-BR, zh-CN, zh-TW) retired on 2026-05-15. Repository now ships English (README.md) + Japanese (README.ja.md) only. Decision driven by traffic data: 30-day GitHub traffic showed 11 unique human viewers across all README versions combined, while 106 unique cloners (and downstream LLM crawler fetches) drove the actual interaction. LLM crawlers (ChatGPT, Qwen, Gemini observed 2026-05) reliably translate the English source on demand, making the mirrors performative rather than functional. Glossary, CODEMAPS, and ADR-0012 updated to reflect the two-language convention. Prior content remains in git history for restoration if audience evidence changes.