Tune local model prompt and output cleanup

[Jam-Cai]

Summary

• Increases local completion token budgets and prompt context window for better short-form continuation quality.
• Tightens local-model prompt instructions around tone, casing, indentation, punctuation, and code continuation.
• Expands backend-neutral output cleanup for MLX/HuggingFace chat-template control tokens.
• Updates prompt, normalizer, request factory, and value-model tests for the new contracts.

Validation

• swiftlint lint --quiet tabby/Models/SuggestionModels.swift tabby/Support/LlamaPromptRenderer.swift tabby/Support/SuggestionTextNormalizer.swift tabbyTests/LlamaPromptRendererTests.swift tabbyTests/ModelAndPresentationValueTests.swift tabbyTests/SuggestionRequestFactoryTests.swift tabbyTests/SuggestionTextNormalizerTests.swift
• xcodebuild -project tabby.xcodeproj -scheme tabby -destination 'platform=macOS' CODE_SIGNING_ALLOWED=NO build
• xcodebuild -project tabby.xcodeproj -scheme tabby -destination 'platform=macOS' CODE_SIGNING_ALLOWED=NO build-for-testing

Greptile Summary

This PR tunes Cotabby's local-model completion pipeline: token budgets are raised from ~1.5× to ~2× the upper word bound, the context window is widened, two new prompt instructions are added for tone/code fidelity, and backend-specific control-token cleanup is consolidated into a new stripKnownControlTokens helper that handles five additional template markers. All affected tests are updated to match the new contracts.

• Token budget and context window increases (SuggestionModels.swift): maxPredictionTokens goes from 8 → 16, maxPrefixCharacters from 1000 → 2000, and each preset's suggestedPredictionTokenBudget is roughly doubled to accommodate modern subword tokenizers.
• Prompt instruction additions (LlamaPromptRenderer.swift): Two new bullets instruct the model to match the user's language/tone/casing and to preserve code symbols, with corresponding test assertions.
• Control-token cleanup expansion (SuggestionTextNormalizer.swift): Replaces the two hardcoded replacingOccurrences calls with a unified helper that strips five globally-safe tokens everywhere and four ambiguous tokens (<s>, </s>, [INST], [/INST]) only at response boundaries.

Confidence Score: 5/5

Safe to merge — all changes are confined to pure helper logic and their tests, with no side-effectful service boundaries touched.

Every changed file is either a pure value type, a stateless helper, or a test. The token-budget and context-window increases are well-commented and consistently reflected in tests. The new stripKnownControlTokens helper correctly handles the most common template-leak patterns, and the boundary-only treatment of /</s> directly addresses the HTML-strikethrough concern from the previous review thread.

No files require special attention; the only nuance is in SuggestionTextNormalizer.swift's boundary-strip loop ordering, which is a latent edge case rather than an active defect.

Important Files Changed

Filename	Overview
Cotabby/Support/SuggestionTextNormalizer.swift	Extracted control-token stripping into stripKnownControlTokens; expands the strip-everywhere list with five new tokens and adds boundary-only handling for <s>, </s>, [INST], [/INST]. Single-pass loop ordering can miss a <s> exposed after stripping a leading </s>.
CotabbyTests/SuggestionTextNormalizerTests.swift	Adds a combined MLX/HuggingFace token test; existing tests all pass. Four of the five newly added tokens have no isolated test cases.
Cotabby/Models/SuggestionModels.swift	Token budgets raised from ~1.5× to ~2× per preset; maxPredictionTokens doubled to 16 and maxPrefixCharacters doubled to 2000. Well-commented and consistently tested.
Cotabby/Support/LlamaPromptRenderer.swift	Two new instruction bullets added for tone/casing/indentation matching and code continuation guidance. Tests updated to assert on the new strings.
CotabbyTests/LlamaPromptRendererTests.swift	Two new assertions verify the added prompt instruction lines are present in the rendered output.
CotabbyTests/ModelAndPresentationValueTests.swift	Token budget assertions updated to match the new ~2× multiplier values. Straightforward numeric updates.
CotabbyTests/SuggestionRequestFactoryTests.swift	Updated the maxPredictionTokens assertion to 40 and extracted a long inline string into a named variable for readability.

_{Reviews (3): Last reviewed commit: "Address review: fix token-budget doc and..." | Re-trigger Greptile}

[FuJacob]

was this for support mlx??