Make uv resolve gpcca-fast without compiling PETSc

[Marius1311]

Problem

Any bare uv command in this repo (uv sync, uv run, uv lock) tried to compile PETSc from source and failed. Root cause:

• There is no committed uv.lock, so uv locks the entire dependency universe — including the opt-in gpcca-fast group (cellrank[petsc], slepc).
• petsc4py/slepc4py ship no wheels and compute their metadata dynamically at build time, so uv had to build PETSc just to resolve — even with --no-default-groups (that only affects install, not locking).
• On machines without a system BLAS on the linker path (isolated CI/uv build envs), PETSc's configure then fails with Could not find a functional BLAS. (Verified from the actual configure.log.)

Net effect: focused runs documented in AGENTS.md like uv run pytest tests/model were broken out of the box.

Fix

Add [[tool.uv.dependency-metadata]] entries that mirror the deps petsc4py/slepc4py inject at build time (petsc4py → petsc, slepc4py → petsc4py + slepc; the two library packages have no Python deps). uv now resolves and locks these from static metadata without ever building — resolves in ~5s. PETSc/SLEPc compile only when the gpcca-fast group is actually installed.

Also corrected the group's source-build comment, which was copied from CellRank's docs and only works where a system BLAS already exists; it now documents the pre-built conda-forge/pixi path as the recommended install, with a shared-BLAS source build as the fallback.

Verification

• uv lock → Resolved 302 packages in 4.80s, no build.
• uv sync --extra test → installs in ~11s, petsc4py correctly absent from .venv.
• uv run pytest tests/test_utils.py → 19 passed.
• hatch test tests/test_utils.py (unchanged path) → 19 passed.

Notes

• uv.lock stays gitignored (library convention; hatch manages the test matrix and doesn't consume it). The static metadata is what unblocks resolution, so no lockfile is required.
• Keep the dependency-metadata entries in sync with the gpcca-fast group.
• The gpcca-fast runtime backend still builds from source only where a real BLAS is available; conda-forge binaries are the robust route (documented). Unchanged by this PR.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.37%. Comparing base (54aff95) to head (8a7de92).

Additional details and impacted files

@@           Coverage Diff           @@
##             main      #86   +/-   ##
=======================================
  Coverage   86.37%   86.37%           
=======================================
  Files          13       13           
  Lines        1387     1387           
=======================================
  Hits         1198     1198           
  Misses        189      189           

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.