refactor(agent): decouple nvm bootstrap from agent Run/InstallMCP commands

[zpzjzj]

Summary

Splits the nvm/Node bootstrap and the agent invocation into two distinct rt.Exec calls so failures can be cleanly attributed and the agent's stderr buffer stays free of bootstrap noise.

Fixes #72

Problem

nodeRuntimeCommandWithGuard (internal/agent/node_install.go) packed the bootstrap and the agent's real command (claude -p, codex exec --json, claude mcp add, codex mcp add) into a single multi-line shell script handed to one rt.Exec call. Consequences:

• Bootstrap failures (curl timeout to the nvm installer, sha256 mismatch, nvm install errors against an offline mirror) surfaced as claude-code run failed: ... / codex run failed: .... Hard to triage without reading the wrapped script.
• Bootstrap stderr was concatenated into result.Stderr, which is then scanned by providerAuthFailureSignal / providerRateLimitSignal — widening their false-positive surface.
• Bootstrap stdout could prepend stdout.json artifacts.
• Bootstrap re-fired silently at Run time when Install failed or wasn't called, masking what should be a clear "Install was never called" failure.

Fix

New ensureNodeRuntime(ctx, rt, cliBin, opts) helper runs only the bootstrap as its own rt.Exec. The script's first line short-circuits with exit 0 when command -v <cli> succeeds, so the happy path is one extra command -v check with no output.

Call site (before)	After
claude_code.go:127 nodeRuntimeCommandWithGuard("claude", ...) (Run)	ensureNodeRuntime → Exec
claude_code.go:72 nodeRuntimeCommandWithGuard("claude", ...) (MCP)	InstallMCP calls ensureNodeRuntime once before the per-server loop
codex.go:236 (Run)	ensureNodeRuntime → Exec
codex.go:136, :171 (MCP stdio / http+bridge)	wrap-once in InstallMCP

nodeRuntimeCommandWithGuard and nodeRuntimeCommand are removed — no remaining production callers. qodercli is untouched (it never used the wrapper).

Errors from ensureNodeRuntime wrap with node bootstrap failed: ... so they're trivially distinguishable from agent run failures in result handling and signal detectors. Bootstrap stderr no longer reaches result.Stderr of the agent Exec.

A follow-up commit (de46509) preserves the existing convention of returning a non-nil SessionResult{Engine, ExitCode: 1, DurationMs} on bootstrap failure, so the evaluator still records telemetry for failed cases.

Test plan

• go build ./...
• go test ./... (full suite green)
• make verify (gofmt + vet + revive + golangci-lint, 0 issues)
• New tests in internal/agent/node_install_test.go: script shape (CLI short-circuit at top), Exec-error wrapping, non-zero-exit wrapping with stderr surfacing, empty-cliBin guard
• Updated fake runtimes (claudeCodeTestRuntime, codexTestRuntime) recognize bootstrap commands by content sniffing and short-circuit them so existing lastCommand / execResult assertions still observe the agent command
• Manual smoke: run an eval with claude-code and verify trace shows two distinct rt.Exec calls (bootstrap + run) and that stdout.json contains only stream-json from claude
• Negative smoke: set NVM_SOURCE to an unreachable host; failure should report node bootstrap failed: rather than claude-code run failed:

🤖 Generated with Claude Code