#21998: fix(models): prioritize exact model-id match over fuzzy scoring (#21942)

## Summary - **Bug**: Model resolver fuzzy-matches `gemini-3.1-pro-preview` to wrong model (`gemini-3-flash-preview` or `gemini-3.1-pro`) - **Root cause**: `resolveFuzzy` in `model-selection.ts` runs fuzzy scoring without first checking for exact model ID matches - **Fix**: Add exact model-id match check before fuzzy scoring in `resolveFuzzy` Fixes #21942 ## Problem The `resolveFuzzy` function in `src/auto-reply/reply/model-selection.ts` scores all candidate models using weighted fuzzy matching (Levenshtein distance, substring matching, variant token penalties). When model names are similar (e.g. `gemini-3.1-pro-preview` vs `gemini-3-flash-preview` vs `gemini-3.1-pro`), the scoring can pick the wrong candidate because: 1. The `FUZZY_VARIANT_TOKENS` list includes "preview", and variant token penalties can reduce the score of the correct model 2. Levenshtein distance between similar model names may favor a shorter/different model **Before fix:** Input: `"gemini-3.1-pro-preview"` Output: resolves to `gemini-3-flash-preview` or `gemini-3.1-pro` (wrong) ## Changes - `src/auto-reply/reply/model-selection.ts` — add exact model-id match check at the top of `resolveFuzzy`, before fuzzy scoring - `src/auto-reply/reply/model-selection.test.ts` — add regression test for exact match priority **After fix:** Input: `"gemini-3.1-pro-preview"` Output: resolves to `gemini-3.1-pro-preview` (correct exact match) ## Test plan - [x] New test: exact model ID resolves correctly over fuzzy candidates - [x] All existing model-selection tests pass - [x] Lint passes ## Effect on User Experience **Before:** Using `/model gemini-3.1-pro-preview` could switch to the wrong model (e.g. `gemini-3-flash-preview`), causing unexpected behavior and confusion. **After:** Exact model ID matches always take priority, so `/model gemini-3.1-pro-preview` correctly selects that exact model. <!-- greptile_comment --> <h3>Greptile Summary</h3> This PR fixes a model resolution bug where similar model names like `gemini-3.1-pro-preview` would incorrectly resolve to wrong candidates (`gemini-3-flash-preview` or `gemini-3.1-pro`) due to fuzzy scoring with variant token penalties. **Key changes:** - Added exact model-id match check before fuzzy scoring in `resolveFuzzy` function (src/auto-reply/reply/model-selection.ts:488-492) - Added regression test to verify exact matches take priority over fuzzy matches - Fixed telegram model-buttons parser to allow colons in provider names (e.g. `ollama:cloud`) **How it works:** The fix adds a simple exact match check that runs before the complex fuzzy scoring logic. When a user specifies a model like `gemini-3.1-pro-preview`, the resolver now checks if any candidate's model ID exactly matches the input (case-insensitive) before attempting fuzzy matching with Levenshtein distance and variant token penalties. The telegram fix updates the regex pattern for parsing `mdl_list_{provider}_{page}` callback data to allow colon characters in provider names, preventing spurious LLM runs when using providers like `ollama:cloud`. <h3>Confidence Score: 5/5</h3> - This PR is safe to merge with minimal risk - The changes are minimal, well-tested, and directly address the reported bug. The exact match check is a simple, defensive addition that runs before fuzzy scoring, so it cannot break existing behavior - it only prevents incorrect fuzzy matches. The test coverage verifies the fix works as intended. The telegram regex change is also straightforward and properly tested. - No files require special attention _{Last reviewed commit: 611829d} <!-- greptile_other_comments_section --> <!-- /greptile_comment -->