MCP bridge flicker stopgap: mcprestart/mcprestart-log (shipped) + McpDoctorUseCase pointer (atlas, parked)

[Data-Wise]

Summary

Ships a zero-engineering stopgap for the MCP bridge flicker/degradation
pattern (documented failure mode: 4-min timeout stall, or instant
"tool not found" death — fix is always quit+relaunch Claude Desktop,
never incremental retry).

Branch: feature/mcprestart-stopgap (pushed, ready for PR)

What shipped

• mcprestart — one command, quit+relaunch Claude Desktop
• mcprestart-log — tail of recent restarts + this-month count

Both added to zsh/functions.zsh (Category 11). Live-verified via
zsh -ic before commit — zero syntax errors, both resolve correctly
in an interactive shell.

Context / spec pointer

Full design context lives in ADR-001 (drafted this session,
docs-standards repo — not yet committed as a standalone file, notes
below are the complete spec):

Decision history (this session):

1. Investigated whether craft, atlas, or a new repo should own
   MCP/config/MLX drift-audit tooling. Found atlas/src/use-cases/registry/DoctorUseCase.js
   already implements the exact report-first + --fix-flag pattern
   for the Project Settings Contract (.STATUS/CLAUDE.md/.obs/sync.yml).
   Found craft/governance/ already implements a fixture-tested
   (good/bad pairs) policy-as-code engine, scoped to skills only.
   Found workspace-auditor is a paused (~10%) stub whose own CLAUDE.md
   already flagged this exact atlas-vs-new-repo question — recommend
   formal retirement.
2. Decision: atlas hosts a new McpDoctorUseCase (sibling to
   DoctorUseCase), same execute()/fix({write}) contract.
3. Grilled the plan hard before committing engineering time:
   • Confirmed flicker frequency is real: 5+/month (measured
     recall, not yet logged — that's what mcprestart-log now fixes).
   • Dashboard-tile idea was floated then explicitly dropped — user
     caught it as "avoiding a decision" rather than a real requirement.
     CLI-only, no dashboard mechanism.
   • Fixture format gap identified: MCP flicker is a sequence
     (success → timeout → timeout), not a static snapshot — craft's
     existing good/bad fixture pattern doesn't cover this as-is.
     Decision: design a sequence-aware fixture format before writing
     the check — not deferred, not skipped.
   • Priority check: McpDoctorUseCase is confirmed not this
     week's actual priority — genuinely parked, not abandoned.

Cross-repo

Pointer issue filed in atlas with the full McpDoctorUseCase spec
(same content, atlas-side framing) — see: [link added after atlas issue is created]

Follow-up (this repo)

• Open PR for feature/mcprestart-stopgap → dev
• Check-in in 2-4 weeks: run mcprestart-log, compare real count
  against the "5+/month" estimate — this number is the actual
  trigger for whether McpDoctorUseCase jumps the queue

[Data-Wise]

Cross-link: atlas-side spec filed as Data-Wise/atlas#70.