feat(bench-kit): zone-aware update + agent-tool profile in init

[psmyrdek]

Implements the two remaining bench-kit CLI pieces (points 1–2 of the current plan; design contract: DESIGN.md in 10x-benchmark-webinar / template README zone contracts).

Stacked on #30 (init copies workflows) — merge that first, then retarget/merge this into master.

init: agent-tool recognition

• Resolution chain: --tool flag → manifest tool on repair → interactive @clack pick pre-filled by lib/tool-detect.ts marker detection → detection result → claude-code (also the --yes path).
• Template skills are materialized under the chosen profile's skill root (.claude/skills/, .agents/skills/, .cursor/skills/, …) instead of hard-coding the template's location; no empty .claude/ shell is left behind.
• The chosen tool lands in .bench-kit/instance.json so update knows where to propose skill diffs. Repair preserves the manifest (tool, initializedAt, detected base repo).

update: zone-aware template upgrade

Mirrors the manual adoption procedure done four times (0.3.0, 0.4.0, 0.4.1, 0.4.2):

Zone	Behavior
.bench-kit/	replaced wholesale via staged swap; manifest survives, version-bumped (updatedAt)
.github/workflows/	synced from the new template
skills (per manifest tool)	synced into the working tree as an uncommitted proposal — company reviews git diff; company-only skills never deleted
tasks/, evaluation-pool/, bench.config.yaml	untouched

• Clean-worktree gate (dirty_worktree) so the proposal diff stays reviewable; up to date no-op when versions match; not_an_instance guard.
• Closing hint points at bench validate, which owns schema-compatibility checking per the design.
• Template skill source auto-detects the planned .agents/skills/ migration (falls back to .claude/skills/).

Testing

• 22 unit tests in tests/bench-kit-command.test.ts (init tool placement, repair preservation, all update zones + guards); full suite 503 green, typecheck + lint clean.
• Also fixed a latent race in the test helper captureStreams (restore ran after the next capture's patch, leaking output to real stdout).
• E2E against the real template: init --template-version v0.3.0 --tool codex → skills under .agents/skills/; customized a template skill + added a company skill → update to 0.4.2: runtime replaced (new leaderboard files in), skill customization shown as reviewable diff, company skill preserved, manifest bumped, dirty guard verified.

🤖 Generated with Claude Code