[patty-chow]

Summary

Automatically prepare bundled vector search during both Codespaces profiles' startup, without automatically pulling Git changes or changing attendee exercises.

• codespace:start first starts/seeds the core planner, then prepares both local sites for search. Core-ready and search-ready/unavailable messages are distinct.
• Valid caches are verified by the actual model manifest's full SHA-256 and size plus ollama show; cached startup makes no host model-registry request and never starts the downloader.
• Missing/mismatched cache uses the controlled download path with bounded signal-resistant downloader cleanup.
• Shared checkout lock prevents concurrent automatic/manual preparation. Opt out with WORKSHOP_AUTO_SEARCH=0.
• Optional search failures leave core work usable; cancellations and cleanup danger remain nonzero. Normal workspace:start remains the core-only fallback and does not pull Git or contact Azure.
• Updated setup, presenter/local and acceptance documentation to distinguish automatic Codespaces startup from the manual fallback.

Evidence so far

• Local tests: 181 passed, one existing Linux-only skip; 75 docs links checked.
• Owned local empty-model-volume first startup: 55.72s including model download (Docker images were cached).
• Cached local automatic start: 13.49s with host registry fetch forbidden; zero external fetches and downloader start/stop timestamps unchanged.
• Actual search integration 1/1 and database integration 2/2 pass. Injected registry failure kept actual core health/state 200 with 216 sessions; opt-out/manual retry passed.
• All owned local services stopped, volumes retained, unrelated services untouched.

Completed Codespaces and review gates

• Fresh Attendee Codespace ran the actual postStart hook without manual commands: core HTTP200 at ~4m30, search ready at ~6m39 including first runtime/model downloads; both sites loaded 216 bundled vectors with zero document inference.
• Actual stop/Shutdown/start: cached search phase 12.951s, total VM/core/search resume 83.792s. Downloader ID/timestamps unchanged; supplementary fetch-forbidden startup proved zero host registry fetch calls.
• Browser untouched TODO1 plus search ready and real 0.747/0.724 results; real DB integration 2/2, search 1/1. Controlled registry-failure injection preserved core200; opt-out/recovery passed.
• Review caught a retry-after-cleanup-failure gap; fixed at 92bc4a9 with pre-inference downloader inventory and a real-lock two-invocation regression. Reviewer approved the correction.
• Targeted actual Codespace guard proof: live running downloader refused before startup/cache/fetch/index; other service states unchanged; scoped recovery then cached retry succeeded with zero fetches and unchanged downloader timestamps (10.03s search phase).
• Final Codespace root tests: 195/195, docs75. Final documentation adds the measured evidence with exact revision scope; no additional runtime changes.
• Environment/exercise-file hashes unchanged. Only the owned test Codespace was used and stopped; data retained. Organizer profile has config/test coverage but was not separately created.

No Azure resources, paid API keys, notebook deletion or automatic Git synchronization.