test: raise non-generated Lua coverage to ≥90% (issue #45)

[levleontiev]

Summary

• Adds targeted unit coverage across 10 spec files; zero runtime code changes
• Raises non_generated coverage from ~87.6% → 91.62%
• All per-module coverage now ≥ 90%

Coverage delta (key modules)

Module	Before	After
decision_api.lua	75.43%	91.12%
bundle_loader.lua	86.73%	90.42%
cost_budget.lua	89.15%	≥90%
cost_extractor.lua	89.74%	≥90%
overall non-generated	~87.6%	91.62%

Approach

All new tests use targeted direct coverage: describe blocks appended after the existing Gherkin runner call. For decision_api_spec and bundle_loader_spec, internal helper functions are reached via debug.getupvalue — a standard LuaJIT testing pattern that avoids exposing internals in the production API.

Test plan

• busted spec/unit/ → 631 successes / 0 failures / 0 errors
• busted spec/integration/ → 65 successes / 0 failures / 0 errors
• pytest tests/e2e -v → 44 passed
• bin/ci/run_lua_coverage.sh → 696 successes, non-generated coverage: 91.62%

Closes #45

🤖 Generated with Claude Code