Implements a validation + repair pipeline for tool calls from models
with known tool-calling quirks (MiniMax M2.5, GLM-5, Grok 4).

## Files

- src/agents/tool-call-validator.ts — validates tool calls, reports
  typed issues, determines repairability via fuzzy matching
- src/agents/tool-call-repair.ts — applies 7 ordered repair strategies
  to produce a best-effort corrected call
- src/agents/tool-call-validator.test.ts — 50 unit tests
- src/agents/tool-call-repair.test.ts — 98 unit tests

## Validator

Reports typed issues:
  - unknown_tool (fuzzy-aware repairability check)
  - invalid_id
  - empty_args
  - malformed_json
  - missing_required
  - wrong_type (notes coercibility)
  - extra_params

## Repair strategies (in order)

1. JSON fix-up: trailing commas, single quotes, unquoted keys,
   missing braces/brackets, BOM removal
2. Type coercion: string->boolean, string->number, scalar->array
3. Tool name fuzzy matching: Levenshtein <= 3 or case-insensitive
4. Parameter name normalisation: camelCase <-> snake_case
5. Missing required params relocation from extra fields
6. Tool call ID generation/sanitisation
7. Provider-specific fixes:
   - MiniMax: unwrap object/string wrappers, double-encoded JSON,
     malformed strings via fixJson
   - GLM-5/ZhipuAI: unwrap parameters/function_arguments wrappers
   - Grok 4: same patterns as MiniMax
   - Generic (unknown provider): heuristic single-key wrapper detection

## Test results

148/148 passing, 0 skipped
TypeScript: clean (tsc --noEmit exit 0)
ESLint: 0 errors, 0 warnings

Levenshtein distance implemented from scratch, no external deps.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>