[sanskarIN]

User-visible issue

VS Code cursor columns are UTF-16 code-unit offsets, so a manual or automatic inline completion can be requested at a position between the halves of an emoji or non-BMP letter. Although context extraction trims orphaned halves from prompts, inserting completion text at that cursor could split the actual editor character. Separately, prefix-echo and suffix-overlap filtering treated surrogate halves as non-identifier punctuation and could truncate genuine non-BMP identifiers.

Implementation

• Add a zero-allocation pure isSafeInsertionBoundary check: only valid UTF-16 line offsets outside surrogate pairs are eligible. Apply before any context collection, Ollama request, or manual/automatic response path.
• Make prefix echo boundaries and suffix overlap guards compare whole Unicode scalar values rather than individual UTF-16 units, preserving true astral Unicode identifiers while still stripping unambiguous repeated code at punctuation boundaries.
• Add deterministic regression cases across emoji, ZWJ sequences, Deseret letters, Hindi source, mixed ASCII/astral identifiers, fractional/invalid cursor positions, and safe suffix echoes.
• Add a manual physical VS Code QA checklist at docs/qa/unicode-inline-boundary-acceptance.md, explicitly NOT RUN.

Gates and limitations

Full cross-platform CI, unit tests, Linux mocked editor-host integration, security audit and isolated VSIX candidate installation/upgrade must succeed before merge. There are no new network calls, model downloads, trained weights, repository indexes, publisher-ID changes, GitHub releases or registry publications. Physical editor and real Ollama model performance remain unverified.