runner: Shutdown closes scoped MCP registries of wedged runs (audit #16)

[dennisonbertram]

Intent

Make Runner.Shutdown iterate r.runs and close scoped MCP registries so a wedged run's MCP subprocesses are torn down even when execute() never reaches a terminal-cleanup path.

Plan reference

• Plan: docs/plans/2026-06-24-harness-reliability-plan.md — slice T05
• Umbrella epic: epic: harnessd long-session reliability hardening (15 TDD slices, audit 2026-06-24) #644
• Audit finding: JSONL rollout recorder for session replay, fork, and audit #16

Source finding

ScopedMCP subprocesses are not killed if a run is wedged at Shutdown (runner.go:3067-3091, mcp.go:662-665). closeScopedMCP only runs from the four terminal-cleanup paths. If execute() is blocked (e.g. a deadlock, or an MCP CallTool ignoring ctx), Runner.Shutdown only cancels cancelFuncs and does not iterate r.runs to close scoped registries.

Scope

• Allowed files (test): internal/harness/runner_shutdown_mcp_test.go (new)
• Allowed files (impl): internal/harness/runner.go, internal/harness/scoped_mcp.go
• Dependencies: none
• Complexity: low — Risk: low (idempotent Close)

Acceptance criteria

Use slice T05 in the plan as detailed acceptance criteria. Summary:

• Failing test written first, compiles and fails on main.
• Stub a wedged scoped MCP registry whose ClientManager.Close blocks until called. Start a run, wedge its execute goroutine by injecting a tool that blocks on a channel. Call Runner.Shutdown; assert ClientManager.Close was called and the underlying MCP subprocess (a fake os/exec recording KILL via cmd.Process.Kill) was signalled.
• Impl: closeScopedMCP is already idempotent — set state.scopedMCPRegistry = nil at the end so re-entry is a no-op. In Shutdown, snapshot state.scopedMCPRegistry for every live *runState, then _ = reg.Close(shutdownCtx) after cancelling cancelFuncs. Add defer r.closeScopedMCP(runID) inside execute() so a panic still tears down MCP.
• go test ./internal/harness/... -race green; ./scripts/test-regression.sh green.

Guardrails

• Strict TDD; failing regression test must compile and fail on main first.
• Worktree flow; own branch, merge via verify-and-merge.sh.
• Update .context/harness-reliability/tracker.md on pick-up and merge.
• Deterministic tests only — fake MCP registry + fake exec.
• Append engineering-log entry on merge.