test: tool call regression suite for compressed models (closes #96)

tests/tool_call_regression.py:
- 10 test cases covering 5 hermes tools: read_file, web_search, terminal,
  execute_code, delegate_task
- Schema validation (OpenAI-compatible tool call format)
- Argument validation (correct tool + expected args)
- Parallel tool calling test (multiple tools in one response)
- Dry-run mode for CI (schema validation without server)
- Full server mode with latency tracking
- Markdown report generation with results matrix
- JSON results output for programmatic consumption
- 95% accuracy threshold gate (exit code 1 on failure)

benchmarks/tool-call-regression.md:
- Results template with model/preset matrix
- Tool coverage tracking table

.gitea/workflows/smoke.yml:
- Added dry-run tool call schema validation step

benchmarks/tool-call-regression.md

@@ -0,0 +1,32 @@
+# Tool Call Regression Results
+
+**Generated:** 2026-04-16T01:56:48.462512+00:00
+**Model:** dry-run
+**Endpoint:** none
+**KV Type:** none
+
+## Summary
+
+| Metric | Value |
+|--------|-------|
+| Total tests | 10 |
+| Passed | 10 |
+| Failed | 0 |
+| Accuracy | 100.0% |
+| Threshold | 100% |
+| Verdict | PASS |
+
+## Test Matrix
+
+| Test ID | Tool Expected | Tool Called | Schema | Args | Latency | Status |
+|---------|--------------|-------------|--------|------|---------|--------|
+| read_file_basic | read_file | none | OK | OK | 0ms | PASS |
+| read_file_offset | read_file | none | OK | OK | 0ms | PASS |
+| web_search_basic | web_search | none | OK | OK | 0ms | PASS |
+| terminal_basic | terminal | none | OK | OK | 0ms | PASS |
+| terminal_complex | terminal | none | OK | OK | 0ms | PASS |
+| code_exec_basic | execute_code | none | OK | OK | 0ms | PASS |
+| code_exec_complex | execute_code | none | OK | OK | 0ms | PASS |
+| delegate_basic | delegate_task | none | OK | OK | 0ms | PASS |
+| delegate_context | delegate_task | none | OK | OK | 0ms | PASS |
+| parallel_two | read_file | none | OK | OK | 0ms | PASS |