[hmziqagent]

What

Three strands, 10 commits, all gates green (fmt, clippy -D warnings, 524 + 2286 tests).

Harness catalog: 51 -> 91

Two web-survey rounds added 40 missing harnesses, each with an upstream-verified
research doc, catalog entry, concrete adapter, and fixture corpus:

• Nine vendor CLIs (muse-code, qoder-cli, zulu-cli, minimax-code, coco, fx, jules,
  cursor-agent, kilocode-cli)
• Thirteen tier-two surfaces (devin, replit-agent, copilot-vscode, oh-my-opencode,
  oh-my-codex, t3code, orca, bb-app, avante-nvim, codecompanion-nvim, codegeex,
  terminus, qoder-ide)
• Eighteen tier-three indie agents (every-code, claw-code, agentty, command-code,
  eca, freebuff, gitclaw, hax, jcode, omo-ai, omp, openfang, ori, prime-agent,
  reasonix, unreal-agent, vix, zaly)

Partitions: support 26/46/1/7/7/2/2, MCP 34/22/35, plugins 11. Ledger tests now
derive from ENTRIES against named constants plus property assertions.

Fixes surfaced by the audit loop

• opencode v2 config surfaces modeled alongside the v1 era (key renames, cli.json,
  opencode.db, OPENCODE_CONFIG_DIR relocation)
• plugin previews no longer mutate where commit requires approval (plugin.rs)
• cursor evidence probe matches its documented mcp.json path
• nvim-based harness configs isolate under the instance root (XDG_CONFIG_HOME
  relocation, NVIM_APPNAME unset)

Usage-pillar planning

Subplans 14/15/16 (usage core, usage sources, VPS deploy) indexed from the master
plan, plus paste-ready companion plans under docs/plans/export/ for the llmproxy
and toride repos. The CodexBar core extraction (99 provider directories, 29 mapped
provider rows covering 43 harness ids) is documented in subplan 14.

Notes

• letta-code was found already catalogued mid-run, so the real delta was 40, not 41.
• Facts upstream sources do not settle are marked unverified in each research doc;
  nothing was fabricated to fill gaps.