docs(TODO.impl): closeouts - lite2 negative, soup negative, framing pre-registered

[ronaldtse]

Closes out every axis that resolved today, each with its measured number in the file:

• 04 (lite-2.0): NEGATIVE — 7.1402 vs run-009's 5.78. Same recipe, same canonical labels, same teacher; the single variable (layer-drop init from the trained 2.1 instead of generic byt5-small) made the student 1.36pp worse. Generic pretraining keeps layers redundant; task adaptation co-specializes them. The depth axis reads 2-of-3 negative; the lite tier is unchanged.
• 02 (teacher selection): soup postscript — NEGATIVE, 2.4188. The r6+r7 50/50 average is a functional model (same-basin confirmed) but strictly between its parents on every domain; r7 remains the supervision choice.
• 01 (speculative): closed with the framing finding; the tier is pulled from the playground and the runtime keeps measurement infrastructure (see 11).
• 11: pre-registered — static activation scales as the structural framing fix, gates defined before measurement. Python-ORT measurement in: framing equal (480/480) for fp32, dynamic AND static int8 — the runtime-measured instability is onnxruntime-node-specific; static scales are the portable fix candidate (also +8% faster than dynamic on CPU).
• README statuses synced; memo updated with the speculative closeout.