Retry empty final answers

[GTC2080]

Summary

Fixes #3290.

• Treat a no-tool-call model turn with no visible answer text as incomplete instead of ending the user turn.
• Add a bounded retry that asks the model to continue and produce a visible answer, while preserving the existing tool-call and final-readiness gates.
• Add regression coverage for reasoning-only replies and repeated empty replies, plus a microbenchmark for the visible-answer check.
• Update the statusline cache fixture so usage seeding still uses a valid visible assistant reply.

Validation

• go test ./internal/agent -run "TestRun(RetriesReasoningOnlyFinalAnswer|StopsAfterRepeatedEmptyFinalAnswers)$" -count=1 (failed before the fix, passes after)
• go test ./internal/agent -run "TestRun(RetriesReasoningOnlyFinalAnswer|StopsAfterRepeatedEmptyFinalAnswers)$" -bench BenchmarkHasVisibleFinalAnswer -benchmem -count=5
  • normal visible answer: ~1.7-1.9 ns/op, 0 allocs/op
  • 256 leading spaces: ~90-91 ns/op, 0 allocs/op
  • all-space input: ~250-260 ns/op, 0 allocs/op
• go test ./internal/agent -count=1
• go test ./internal/provider/openai ./internal/control -count=1
• go test ./internal/cli -run TestStatuslineKeepsCacheRatesInPrimaryDataRow -count=1
• go test ./... -p 1 -count=1
• go vet ./...
• go test ./... -p 1 -count=1 from desktop/
• npm run build from desktop/frontend/ after generating local Wails bindings and refreshing local dependencies from the lockfile
• npm run test:todo-visibility from desktop/frontend/ -> 200,000 checks in 0.95 ms
• git diff --check