docs: DeepSeek-V4.1-Flash learnings + TODO.impl work ledger

[ronaldtse]

The DeepSeek-V4.1-Flash technical report, read end-to-end and mapped to our stack. Most of it (KV-cache compression at 552B/1M-context scale) does not apply; five levers and two recipe validations do, and two of them are already measured on our artifacts.

Learnings memo (TODO.qwen-next/10), including our measurements:

• Speculative decoding unparked: the shipped layerdrop-int4/small-2.1-int8 pair is a drafter/verifier; probe acceptance 0.97-1.00, ~8.5-8.9 tokens per verifier pass at K=8. TS implementation in interscript-ts#77.
• Multi-teacher routing tested and NEGATIVE: r6 run over all 1200 SadeedDiac rows, sliced by domain - r7 wins every domain (classical tie, news 3.21 vs 3.52, wiki 2.08 vs 3.29). Combined with the student-side slice (gap on news/wiki, not classical), the residual is student-side.
• lite-rung premise corrected by spec audit: run-009 IS trained (6ep sequence-KD); the untested variable is the layer-drop init source (generic byt5-small vs the trained 2.1). The lite-2.0 run is specced and launched.
• Head-wise Muon externally validated (DS V4.1, GLM-5, Kimi-K3); implemented in ml-models with tests.
• Engram notes corrected: it is a hash-addressed n-gram lookup table, not example retrieval; mHC note carries the single-pass shift.

TODO.impl/: the resulting work ledger, one file per workstream (01-10), priorities P0-P3, acceptance criteria per file.