[ponytail] Needs decision: cuts that remove shipped or documented behaviour #127

Not implemented until the maintainer decides

Each item below removes or degrades something that is shipped, documented in CLAUDE.md / README.md, or the subject of a recorded decision. The audit flagged them as over-engineering; the call is the maintainer's. Reply on this issue with the items to cut and they become sub-issues.

• yagni: rollback machinery (src/interfaces/cli/rollback.ts:106, ~380 lines): persisted plan files with a progress marker, 10-file retention, the empty BREAKING_MIGRATIONS refusal, the auto-rollback offer, --rollback / --list-rollbacks / --restore-data. Shipped in 0.4.0 under issue [PRD] engram update --rollback + install-path sanity check #65 / decision Decision: update --rollback: code-only by default or full data restore? #66. The backup directory plus the printed manual recipe would be the whole feature.
• stdlib: the hand-rolled TOML writer for the Codex host (src/interfaces/cli/hosts.ts:356, ~190 lines). Alternative: print the three-line [mcp_servers.engram] snippet for the user to paste. Degrades engram mcp install codex ([PRD] One-command host registration (engram mcp install) #50).
• stdlib: the hand-rolled unified diff used only by engram mcp install --dry-run (src/interfaces/cli/hosts.ts:496, ~75 lines). Alternative: print before and after.
• yagni: the extraction smoke check consolidates real extracted facts into the live database with source='smoke' (src/interfaces/cli/smoke.ts:200, ~60 lines). Alternative: extract and discard. This behaviour is documented in CLAUDE.md.
• delete: shim-script archaeology in the install-path check (src/interfaces/cli/install-path.ts:86, ~55 lines). It exists for npm shims on Windows; nobody has run it on Windows.
• yagni: the --json envelope for engram setup (src/interfaces/cli/setup.ts:108, ~50 lines).
• delete: the --inline-token flag (src/interfaces/cli/hosts.ts:271, ~20 lines), documented under decision Decision: mcp install: write ENGRAM_MCP_TOKEN into host configs or reference the env var? #52.
• shrink: import "newer-wins" arbitration in src/interfaces/cli/transfer.ts:318 (~120 lines). Alternative: an idempotent re-import skips an existing id.
• yagni: Rust and Go parsers in src/graph/file-indexer.ts:873 (~60 lines). TS / JS / Python cover every caller in this repo.
• yagni: the Hermes plugin stdio transport (interfaces/hermes-plugin/provider.py, mcp_client.py, cli.py, config_schema.py, ~889 lines). HTTP is the default, deploy.sh never copies cli.py / config_schema.py, and engram mcp already bridges stdio to HTTP. Issue [hermes-plugin] provider.py replaces a client whose reader thread died without stop() — the still-running node child (model loaded) is orphaned, an exited one is never wait()ed #105 is a bug in this code.
• yagni: the three bash installers (scripts/install-daemon.sh, install-mcp-daemon.sh, install-visualizer.sh) re-copy need / resolve_node / build_engram / ensure_env_file / health-probe / log-tail (~180 lines). Alternative: one sourced scripts/lib/common.sh.
• yagni: the three PowerShell installers are the same file modulo label, port and health path (~170 lines). Alternative: one shared .psm1. Nobody can run these here; parse-check only.
• native: scripts/run-mcp-daemon.ps1 and run-visualizer.ps1 hand-roll a restart loop, rate window and pid files that Task Scheduler's -RestartCount / -RestartInterval already provide (~155 lines). Same Windows caveat.
• delete: .github/workflows/dependabot-native-label.yml (44 lines) attaches one label; the dependabot group and the CI preflight gate already isolate native modules.