Govern completion length by token budget only

[FuJacob]

Summary

Completion length was enforced two ways: an in-prompt word-range cue ("Return only the next 7 to 12 words.") and a token cap. This makes the token cap the single source of truth — the word-range cue is removed from both the local-model (LlamaPromptRenderer) and Apple Intelligence (FoundationModelPromptRenderer) prompts, and suggestedPredictionTokenBudget is bumped ~50% (11/18/30 → 17/27/45) so the cap has room to stop on a natural boundary instead of hard-truncating mid-thought. Both engines already read the same request.maxPredictionTokens, so the cap stays in sync across them.

Validation

xcodebuild -project Cotabby.xcodeproj -scheme Cotabby -destination 'platform=macOS' build-for-testing
# ** TEST BUILD SUCCEEDED **

swiftlint lint --quiet
# clean for changed files

Note: local xcodebuild test could not run the app-hosted bundle due to a Team ID code-signing mismatch on this machine (documented limitation); CI runs with a valid signing identity. Test logic was updated to match the new behavior:

• LlamaPromptRendererTests / PromptPolicyTests / CustomRulesTests now assert the word-range cue is absent from both prompts.
• ModelAndPresentationValueTests updated to the new token budgets (17/27/45).

Linked issues

Risk / rollout notes

• Behavior change to an existing user flow. With no in-prompt target, the word-count presets (3-7, 7-12, 12-20) become ceilings rather than targets. The token budgets allow roughly ~12 / ~20 / ~33 words at the top end (~0.75 words/token), so shorter presets can now overshoot their label. The model still tends to stop at sentence boundaries on its own.
• completionLengthInstruction stays wired through SuggestionRequest and both renderers (Llama via _ =), so re-enabling the in-prompt cue is a one-line revert in each renderer.
• No schema, settings, or pbxproj migrations.

[FuJacob]

Superseded — these changes already shipped to main. This branch was the base of fix/ghost-text-size-stabilization, which merged as #251. Because #251 was a squash merge, GitHub can't see this PR's original commits in main's history, so it stayed 'open' even though the token-budget completion-length change is live (the prompt renderers on main are byte-identical to this branch). Closing to clean up; nothing to merge.