Preserve internal dispatch timings and add AI harness guidance

[anth-volk]

Fixes #2

Summary

• Preserve segment timings from internal Flask dispatch requests on the parent worker operation.
• Keep non-internal nested request timings isolated so unrelated request contexts do not leak into parent operations.
• Add regression tests for both timing propagation paths.
• Add model-agnostic AI harness adapters for agents, Claude, and Copilot.
• Add canonical engineering skills under docs/engineering/skills/, including the PR rule that AI models must run the formatter and linter before the final commit.
• Add Towncrier changelog fragments for the runtime fix and AI harness addition.

Root Cause

Internal dispatch requests created their own request-level timing dictionary. Those request logs are intentionally suppressed, so endpoint segment timings could disappear from the visible modal_worker_dispatch or cloud_run_worker_dispatch operation log.

Validation

• uv run --extra dev ruff format .
• uv run --extra dev ruff check .
• uv run --extra dev --extra all coverage run -m pytest
• uv run --extra dev --extra all coverage report