Kindex AGENTS.md changelog

https://raw.githubusercontent.com/wandercom/kindex/HEAD/AGENTS.md

− Install Python test dependencies (`pip install -e ".[dev,mcp]"`) and Node.js 24 or newer.− The `CI` workflow runs full `pytest` with Python 3.12, Node.js 24, and `.[dev,mcp]` on pull+ Install Python test dependencies (`pip install -e ".[dev,mcp,vectors]"`) and Node.js 24 or newer.+ The `CI` workflow runs full `pytest` with Python 3.12, Node.js 24, and `.[dev,mcp,vectors]` on pull

https://raw.githubusercontent.com/wandercom/kindex/HEAD/AGENTS.md

−    The registry is how MCP clients discover the server, and it stalled at 0.38.0 through six PyPI releases because nothing automated it.− - Generated `.kin` artifacts (`.kin/index.json`, `.kin/code-map.json`): never hand-resolve a git conflict in them. `kin index` auto-registers the `kin merge-kin` structured union merge driver (via `.gitattributes` + local `.git/config`); `kin setup-merge` (re)installs it per clone. Regenerate with `kin index` / `kin export code-map` rather than hand-editing the JSON.+ ## Project knowledge preservation+ Kindex databases hold durable state and may be its only complete copy. Capture+ and update knowledge through the supported Kindex tools in the selected graph.+ Do not delete a Kindex database, including an ignored worktree database, as a+ cache or assume tracked JSON/JSONL files can reconstruct it.+ - Stage and commit matching `.kin/config` and `.kin/index.json` changes with code.+ - Use `kin repo-memory publish` for explicitly selected active public/team+   concepts, decisions, and questions. It is an evidence transport, not a full+   database backup. Keep private knowledge, credentials, and local state out of Git.+ - Preserve both knowledge artifacts if JSON and JSONL coexist. Publish/import+   refuse that layout until their records are explicitly reconciled into one file.+ - Preserve databases and relevant source records during migration or worktree+   cleanup. Report unsupported fields, missing evidence, and unresolved references;+   neither a summary index nor selected exports prove complete recovery.+ The planned canonical JSONL architecture and its unimplemented recovery+ requirements are documented in+ [canonical-knowledge-migration.md](docs/canonical-knowledge-migration.md).+    The registry is how MCP clients discover the server.+ - Generated `.kin` snapshots are not complete canonical knowledge or lossless conflict archives. `kin merge-kin` selects same-ID index conflicts by timestamp (ties keep ours); code-map collisions can keep ours. Preserve both input snapshots and relevant knowledge before regenerating with `kin index` / `kin export code-map` from the selected graph and merged code; an incomplete local graph must not overwrite another branch's knowledge. Do not hand-edit generated snapshots. `kin index` auto-regist