docs(ADR-272): zero-shot tabular foundation model for RuVector (research + plan)

[shaal]

Refs #629.

Research + plan only — no product code. Adds the artifacts needed for @ruvnet to decide whether and how to add a zero-shot tabular foundation model to RuVector (the tabular sibling of ruvector-timesfm).

What's in this PR

• docs/adr/ADR-272-*.md — the decision record: motivation, the license gate + pivot, the four highest-impact surfaces, two-track plan, consequences.
• docs/research/tabfm/PLAN.md — full write-up: TabFM architecture, licensing scorecard, crate-by-crate reimplementation map, and the Track A results.
• docs/research/tabfm/track_a_benchmark.py + track_a_results.json — the reproducible, license-clean benchmark and its raw output.

TL;DR of the findings

• License: TabFM + TabPFN-2.5/2.6/3 weights are non-commercial / non-redistributable → can't ship. TabPFN-v2 (clf+reg) is Apache-2.0 + attribution → shippable.
• Track A (GO): shippable TabPFN-v2 zero-shot on CPU is top-or-tied 5/6 vs tuned XGBoost/HistGBM (mean rank 1.33 vs 2.83), big multiclass win on vehicle. The model we're legally allowed to ship is good enough to own.
• Ranked uses: (1) core candle port, (2) row-compression tabular-row embedder (row→vector, native to the vector DB), (3) inward zero-shot autotuner for HNSW/IVF/quant knobs, (4) Postgres tab_predict() = local AI.PREDICT.

What this PR deliberately does NOT do

No crate scaffold, no weights vendored, no dependency changes. Direction (which surface first, whether to benchmark the gated v2.5/v3 ceiling, whether to eventually train our own weights) is left open for maintainer decision — merge as a research record, or use as the basis for a scoped implementation issue.

Reproducing the benchmark

python -m venv venv --system-site-packages && source venv/bin/activate
pip install tabpfn xgboost         # sklearn/numpy/pandas assumed present
python docs/research/tabfm/track_a_benchmark.py

(macOS: XGBoost needs libomp; the run used the copy shipped in torch/lib. TabPFN-v2 Apache-2.0 weights auto-download; the gated v2.5/v3 columns are skipped unless a Prior Labs license is accepted — internal benchmarking is license-exempt.)