Let AI predictions fill empty prediction slots

[enaboapps]

Closes #1037

What changed

• model.rs: when the AI refinement arrives, its words follow the normal predictions and take the next free slot, so [help, _, _, helpful, helping, helped] becomes [help, helpful, helping, helped, _, _]. There are still at most three AI words.
• worker.rs: each shown word takes the next free slot, so a word dropped because it doesn't match the typed prefix no longer leaves a gap. Labels and inserts share the same index, so accepting a slot still inserts the word shown there.
• Normal predictions keep their order and stay first. They are worked out the same way for the immediate and refined batches, so a refinement never moves a word that's already on screen. While the prediction row is being scanned, the keyboard still holds a refinement back until scanning leaves the row, as before.

Validation

• Independent review on the first head found that the ignored neural_integration_benchmark still assumed AI words were in slots 3-5. That is fixed, and an end-to-end engine test (generated_words_fill_slots_the_instant_words_left) is added. A second independent review on 7f51619 reported no actionable findings.
• New tests: generated_words_take_the_next_free_slot (model merge) and unshown_words_leave_no_gap (worker labels and acceptance); refinement_preserves_the_displayed_acceptance_and_retires_it_on_ack updated for the compacted slot.
• npm run lint, npm test (295 passed), npm run build, cargo fmt --check, cargo clippy --all-targets -D warnings, cargo test (761 passed, 5 ignored): all pass locally on Windows.
• npm run prediction-model failed locally with an EPERM rename in .cache/prediction-neural, a Windows file lock unrelated to this change. CI runs it on clean runners.

🤖 Generated with Claude Code