fix(kiro): normalize version separator mid-identifier

[tonybase]

Summary

normalizeKiroVersion converts a dash version separator into a dot in Kiro backend model IDs (e.g. claude-sonnet-4-5 → claude-sonnet-4.5). The previous implementation only rewrote a trailing <digit>-<digit> pair, so IDs where the version sits mid-identifier were left with a dash and never resolved:

• gpt-5-6-terra stayed gpt-5-6-terra (should be gpt-5.6-terra)
• kimi-k2-7-code stayed kimi-k2-7-code (should be kimi-k2.7-code)

Fix

Rewrite the first dash that sits directly between two digits (<digit>-<digit>). This handles the version wherever it appears, and rewriting only the first such dash keeps trailing build/date segments intact:

• gpt-5-6-terra → gpt-5.6-terra
• kimi-k2-7-code → kimi-k2.7-code
• grok-4-20-0309-reasoning → grok-4.20-0309-reasoning (only the first pair collapses)
• claude-sonnet-4-5 → claude-sonnet-4.5 (unchanged behavior)
• glm-5 / claude-sonnet-5 → unchanged (single digit segment, no digit-dash-digit)

Testing

Added table cases in kiro_executor_test.go covering gpt/kimi/deepseek/grok variants plus the single-segment no-op. go test ./internal/runtime/executor/ passes and go build ./cmd/server is clean.

🤖 Generated with Claude Code