Summary

• Convert pyproject.toml to PEP 621 metadata with the hatchling backend; replace poetry.lock with uv.lock
• Keep main as a project script (uv run main); add taskipy tasks test and benchmark (uv run task <name>), matching the convention in agent-management-api
• benchmark runs artemis_scripts/benchmark.py --runs 3 --output artemis_results.csv
• Fix an fd leak in SqlQuery: connections were never closed and were only reclaimed by cycle GC. On Python 3.14 the new incremental GC reclaims them lazily enough that benchmark runs blow past a 1024 fd limit and crash with EMFILE
• Update README commands accordingly

Transition compatibility (poetry still works)

Falcon rolls out its command change per environment (dev on merge, staging on component release, prod on platform release), while this repo's main is shared by all environments. To avoid breaking staging/prod — which keep running pip3 install poetry && python -m poetry install etc. until their releases ship — the pyproject stays resolvable by poetry 2.x, and poetry.lock is kept alongside uv.lock:

• requires-python = ">=3.8,<4.0" (taskipy caps python at <4.0; poetry needs the bound to resolve)
• psutil relaxed to >=5.9,<8 so taskipy's psutil<7 pin resolves without uv's override-dependencies, which poetry has no equivalent for

Verified the full old command sequence (poetry install, poetry run pytest --benchmark-skip tests/, the benchmark script) against this branch with a fresh pip install poetry (2.4.1). Once the falcon change reaches prod, poetry support (its lock file and the constraints above) can be dropped.

Test plan

• uv run main runs the demo
• uv run task test — 79 passed, 18 skipped
• uv run task benchmark completes 3 runs and writes artemis_results.csv, verified under ulimit -n 1024 on Python 3.14.2
• Old poetry commands verified working against this branch (prod simulation)

Related PRs

• turintech/falcon-optimisation-api#1745 — merge after this PR

🤖 Generated with Claude Code