L6: deliver the "make a todo CLI" journey via a cheatcode (install → use → uninstall) to cover the cheatcode pipeline end-to-end

[ZaxShen]

Goal

Rework the existing "make a todo CLI" L6 chain so the todo CLI is delivered by a cheatcode rather than by SWE directly: bro detects the capability gap → installs a cheatcode that provides the build capability → uses it to ship the todo CLI → uninstalls the cheatcode. If the cheatcode pipeline delivers the CLI end-to-end, SWE inherently works too — one journey covers ~99% of the path and keeps L6 concise.

Why this works in L6

Every L5/L6 step is a fresh claude -p cold start (no claude --resume); bro rehydrates from the trajectory DB. So a cheatcode installed in step N is picked up by step N+1's cold start — install/use/uninstall across chained steps validates hot-load (#660) naturally, with no resume mechanics.

Determinism

• Registry/network stubbed via TMB_CHEATCODE_SEARCH_FIXTURE / TMB_CHEATCODE_VET_FIXTURE.
• Marketplace install/uninstall stubbed via TMB_CHEATCODE_INSTALL_FIXTURE (+ uninstall fixture).
• PreToolUse approval gate satisfied by seeding a cheatcode_approve row via the row's chain_setup.
• Exercise an MCP-kind cheatcode (no prompt-surface edits).

Constraints

• L6 prompt.txt is Human-authored/approved (tests/ hard rule) — bro/SWE may draft, Human signs off.
• Depends on Design: cheatcode uninstall/teardown stage (remove install record + files + agent wiring + audit) #676 (uninstall) + v0.10.0: hot-load / in-session integration — use installed resource without restart, else claude --resume #660 (activate) landing first (tracked in Ship cheatcode lifecycle: cheatcode_uninstall (#676) + cheatcode_activate hot-load (#660), L1-L4 #782).

Acceptance

• The todo-CLI L6 chain ships the CLI via the cheatcode pipeline; assertions cover install (cheatcodes + attachment + audit rows), use (CLI built), uninstall (records reversed + cheatcode_uninstalled audit).

Milestone

v0.10.0