Improve missing-attribute suggestions: Jaro-Winkler similarity, gated containment - #147

What does this implement/fix? Explain your changes.

Improves the quality of the "Did you mean" hints on missing-attribute errors, fixing two defects in how similar member names are selected:

1. 1–2 letter members flooded the hint for long missed names. The substring-containment test admitted any member that appears inside the missed name, and types like QCAlgorithm expose many 1–2 letter members (indicator shortcuts, greek letters), every one of which is a substring of a long name:

   'QCAlgorithm' object has no attribute 'set_account_type' Did you mean: 'cc', 'co', 'a', 'c', 't'?
   

   Containment now only counts for fragments of at least 3 characters, and a candidate contained in the missed name must additionally cover at least half of it.

2. Suffix-extended members were never suggested. The Levenshtein threshold max(2, len/3) rejected members whose real name extends the guess with a suffix, so the member the user most likely meant could be missing from its own hint:

   type object 'BrokerageName' has no attribute 'InteractiveBrokers' Did you mean: 'INTERACTIVE_BROKERS_FIX'?
   

   (INTERACTIVE_BROKERS_BROKERAGE — 11 edits away — was not suggested.) The edit-distance threshold is replaced with Jaro-Winkler similarity (threshold 0.87), a prefix-favoring measure that keeps suffix extensions and, as a bonus, rejects the junk matches short names produced under the distance cutoff (e.g. PII no longer suggests 'min', 'pow', 'sin' alongside 'PI').

Both fleet-reported cases now produce the intended hints against real Lean types:

'QCAlgorithm' object has no attribute 'set_account_type' Did you mean: 'set_account_currency'?
type object 'BrokerageName' has no attribute 'InteractiveBrokers' Did you mean: 'INTERACTIVE_BROKERS_FIX', 'INTERACTIVE_BROKERS_BROKERAGE'?

Algorithm selection. Two candidate replacements were evaluated over real member pools (590 QCAlgorithm members, BrokerageName, System.String, System.Math, DayOfWeek, plus the existing test fixtures): a token-based clause (split on _/camel humps, relate on shared tokens, keep Levenshtein for typos) and Jaro-Winkler. Both pass all 13 graded cases; Jaro-Winkler was chosen because it replaces the edit-distance clause instead of augmenting it, so it also fixes the short-name junk the token variant inherits, with less code:

Threshold margins at 0.87: all intended targets score ≥ 0.90; the trimmed noise sits ≤ 0.851. Substring containment is kept (gated) as a fallback for fragment lookups Jaro-Winkler's match window cannot see (cash → set_cash), scored by coverage so such matches always rank below similarity matches. The per-type candidate cache, per-(type, name) hint memoization, kind filtering and the 5-suggestion cap are unchanged.

Does this close any currently open issues?

No open issue in this repo; addresses the enum/member did-you-mean defects from QuantConnect/Agents#305 (improvement 1).

Any other comments?

Tests: py -3.11 -m pytest --runtime netcore tests → 459 passed, 1 failed (test_module.py::test_explicit_assembly_load, pre-existing environmental failure, reproduced on unmodified master), 23 skipped. dotnet test src/embed_tests/Python.EmbeddingTest.csproj -c Release --filter "FullyQualifiedName!~SetPythonPath" → 971 passed, 0 failed.

Note: #144 moves LevenshteinDistance into the same Util.cs this PR adds the Jaro-Winkler helpers to (and this PR removes that method's ClassBase copy), so whichever lands second has a trivial both-append merge in Util.cs and a deletion conflict in ClassBase.cs.

Checklist

Check all those that are applicable and complete.

• Make sure to include one or more tests for your change
• If an enhancement PR, please create docs and at best an example
• Ensure you have signed the .NET Foundation CLA
• Add yourself to AUTHORS
• Updated the CHANGELOG