Opened a PR for a conversation_metadata race that's been biting self-hosted: the selected repo gets wiped from running conversations (UI flips to "no repo connected"). Root cause is the conversation-start webhook racing the create, reading a stub, and writing the metadata row back with the git fields nulled. Fix makes selected_repository / selected_branch / git_provider write-once (COALESCE), and folds in the duplicate-key insert-race recovery too. Same race as #14476, but the data-loss symptom the open retry PRs (#14579, #14940) don't cover. Green CI + regression tests, reproduced ~40% before / 0 after on a self-hosted SQLite box. Would appreciate a review 🙏 github.com/OpenHands/OpenHands/pull/15015

I siggest you to post this in proj-agent-canvas :-)