Follow-ups from #1738: pin fallback policy, skill-import deadline, harness question ordering #2021

#1738's pair/continue work shipped through #1888, #1950, #1967 and #1990, and its two independent fixes are carried forward in #2020. Three of its changes are still missing on dev and app/v3. Each needs a scope decision before anyone re-implements it:

1. Task and checker model pins vs. inferred fallback. Should an explicit task or checker model pick forbid catalog-inferred fallback, or only choose the starting model? Keeping pins strict needs explicit fallback and provenance rules; allowing rescue means dropping the old pin guards. Today internal/session/task_run.go uses the unfiltered chain, and Use codeaf on two computers: pair with a link, continue a chat on the other #1738's two pin regressions fail on dev, moving to a guessed model. task_standing_model_test.go already distinguishes a proposal's model from a human's standing pick.
2. A deadline for blocked skill imports. cmd/codeaf/chatv3.go still imports skills synchronously before the first message. Use codeaf on two computers: pair with a link, continue a chat on the other #1738 stopped waiting after two seconds but left the import goroutine running. Deciding whether a launch may proceed before the import finishes means settling who owns the shared shelf: bounded concurrency and shelf-close ownership, or a redesign that keeps a complete first-message shelf.
3. Harness question-publication ordering. Use codeaf on two computers: pair with a link, continue a chat on the other #1738's raiseHarnessLane helper is absent. Current offers register their answer lane (internal/session/harness.go), then publish. Any change has to be re-scoped against today's question lifecycle, which app: questions settle everywhere, Hold holds, and late model lists are announced #2015 changed further on app/v3.

🤖 Generated with Claude Code