feat(tracing): infer error ownership from tracebacks

[jshaikScale]

Summary

• automatically infer application vs platform ownership from the innermost relevant traceback frame
• preserve authoritative precedence: explicit category → canonical categorized error → stack inference → known boundary → scoped mapping → unknown
• expose immutable module-prefix and normalized file-root ownership rules that are deterministic and testable without mutable global state
• forward safe classifier provenance to SGP while preserving the existing type, message, and category error fields

Ownership behavior

• Agentex calling user/agent code that raises → application
• user code calling Agentex internals that raise → platform
• ignored third-party/framework frames defer to the nearest relevant owned caller
• conflicting, empty, relative, archive/obfuscated, or otherwise unresolvable stacks → unknown
• sync, async, and ADK context managers classify from the original exception traceback and re-raise the original exception unchanged

The classifier never uses exception messages, exception class names, source lines, raw paths, or stack text as inference signals. Telemetry contains only the stable category source, classifier version, and deterministic non-sensitive rule ID.

Tests

• 198 passed, 2 skipped — full relevant tracing, ADK, and processor suite
• Ruff lint: passed
• Ruff format check: passed
• Pyright: 0 errors, 0 warnings, 0 informations
• focused mypy: Success: no issues found in 2 source files
• git diff --check: passed

Limitations and follow-up

• zip, obfuscated, relative, or otherwise unresolvable frames safely remain unknown unless explicit module ownership is configured
• applications installed beneath shared site-package roots should configure application module prefixes or file roots
• non-Agentex producers still require coordinated adoption in the shared scale-gp-beta SDK/repository