[amrit110]

Version bump for v0.13.0. Once this is merged, I'll tag v0.13.0 on main, which publishes to PyPI through publish.yml.

What's in it

• fix(predict): size predict chunks with a measured predict cost model (#47) #54: predicts no longer run as one-row calls (Fixed 10–30 s stalls on fit and predict across table sizes (recompiles?) #47). A 100-row predict against a 3,500 × 500 context took 14.4 s on an A100 (TabICL) and now takes 0.11 s; on prod's A40s it took about 24 s.
• fix(workers): serialize reaping so a killed worker never looks alive #55: worker reap race. A killed worker could look alive and keep its host-memory reservation.
• fix(predict): classification labels are strings whatever return_proba is (#45) #56: classification labels are always strings (TabICL classification predict returns numeric labels without return_proba, strings with it #45). This changes behaviour for tabicl-v2 without return_proba.
• fix(compile): apply dynamo settings per call -- serving threads saw a limit of 8 (#47) #57: compiled TabICL in serving (Fixed 10–30 s stalls on fit and predict across table sizes (recompiles?) #47). Serving threads saw dynamo's recompile limit of 8, so prod's workers stopped compiling right after warmup.

Changes in this PR

• pyproject.toml and the tabctx entry in uv.lock: 0.12.0 → 0.13.0. Nothing else in the lockfile; a local uv lock only rewrote markers.
• CHANGELOG: the Unreleased entries become [0.13.0] - 2026-10-20.
• README: status line.

Unit tests: 508 pass.

🤖 Generated with Claude Code