bench(tabular): soft-label distillation results + lock the blob format

Full 51-dataset re-run adding gbt<-tabfm soft (TabFM's predict_proba distilled
via proba/classes). Result is modest but exactly what the mechanism predicts:
soft vs hard is 16W/11T/8L (mean +0.003) -- most datasets tie because the
teacher was already confident -- but soft gbt<-tabfm beats the knn5-taught
student on 74% (up from hard's 65%), and on 6 datasets where hard-label
distillation had lost to the cheap knn5 teacher, soft caught back up or passed
it. Soft labels stop discarding TabFM's calibrated edge; what they can't fix is
the representational half of the gap (axis-aligned trees vs a warped boundary),
which is the next student, not the next target. tabarena-full.md updated with
the soft column, head-to-head, and analysis.

ARCHITECTURE.md: point at the now-normative, versioned blob format (RFC 4.1.6,
PSTREE01/PSGBT01) -- a stored student stays servable across upgrades,
snapshots, and forks, and a serving-only module can execute a blob it never
trained.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Signed-off-by: mstrathman <matthew.strathman@gmail.com>

Author: mstrathman

ARCHITECTURE.md

@@ -103,8 +103,11 @@ inherits a foundation model's calibration instead of only its argmax, while
 the holdout is still scored against the true `target` labels. `predict()` dispatches
 a `tree`-runtime model to the native runtime in the same file, which tells a
 single tree (`PSTREE` blob) from a forest (`PSGBT` blob) by magic and needs
-no onnxruntime. Both blob formats are little-endian and implementation-defined
-(RFC §4.2.4), and rigorously bounds-checked on read, because the registry is
+no onnxruntime. Both blob formats are little-endian and normatively specified
+and versioned (RFC §4.1.6, `PSTREE01` / `PSGBT01`), so a stored student stays
+servable across upgrades, snapshots, and forks, and a serving-only module can
+execute a blob it never trained. They are rigorously bounds-checked on read,
+because the registry is
 writable by any SQL caller (RFC §6.2) — a hand-crafted blob is rejected,
 never crashed on. Because the student is native and deterministic, its
 predictions carry the same exact-replay receipt as the stat models.