docs: TODO 11 done (parity corpus + ruby plane runtime + workflow), 12 specs ready (urdu paused)

Author: Ronald Tse

TODO.final/11-neural-parity-goldens.md

@@ -1,6 +1,6 @@
 # 11 — Neural cross-runtime parity goldens
 
-**Status: pending.** Goldens cover maps (7388/7388 parity). Model
+**Status: DONE** (2026-10-07, minus the ts plane port) — golden-v2 released (both plane models, py-reference); RUBY GAINED A PLANE RUNTIME (gap the corpus exposed: plane existed only in py) with cycling-fixture specs, 314 examples green; neural-parity.yml workflow green on both legs. CONTRACT FINDINGS: (a) int8 artifacts are not byte-reproducible across onnxruntime builds/platforms, not even for the reference runtime; (b) K-pass decoding cascades single flips into clustered per-row divergence. Hence two tiers: tight (per-row 2% + corpus 0.5%) and smoke (corpus 2%), byte-exact reserved for deterministic artifacts. REMAINDER: port PlaneModel to interscript-ts (py->ruby pattern), add the ts leg to the workflow.
 inference (IMF seq2seq + plane K-pass) has no cross-runtime corpus:
 each runtime implements decode independently.
 

TODO.final/12-plane-family-expansion.md

@@ -1,6 +1,6 @@
 # 12 — Plane family expansion (Hebrew nikud, Urdu)
 
-**Status: pending (architecture proven).** The plane factorization is
+**Status: SPECS READY, NOT STARTED** — Urdu arm PAUSED (new corpus arriving per owner 2026-10-06).
 language-agnostic; corpora already sit in the data layer (hewiki,
 rababa-hebrew-distilled, urdu corpora).