server /train: validate numbers before examples; the prefix-stability refusal says what to do (Cormac's #20 notes)

[joelteply]

• window/rank/epochs/... are checked first, so a window of 0 is "window must be an integer in
  [16, 8192]" instead of "this window holds -4" from the example preparation that used it.
• a template that rewrites earlier turns (Qwen3 strips earlier turns' ) refuses the
  multi-turn example; the 400 now says to send such a conversation as single-turn examples.

On the 5090 (1.5B): window 0 -> the range message; the #20 mask run is unchanged
(trainable_tokens 1979 of 2177, loss 4.733), as is the text path.

Notes 1 and 3 need no change today. ncorrect is not reported, and AdamW normalizes the gradient scale.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Q4NU4VNiELPQfBpCacDZGc

[joelteply]

APPROVED at fdca596 (Cormac). The validation block moved verbatim above the examples branch, so prepare_examples never sees an unvalidated window. The only other change is the refusal text, which now tells the sender what to do. I read the code and did not build it; the IntelMac is mid-deploy build.

🤖 Generated with Claude Code

https://claude.ai/code/session_01FRQWzgSfo79JtnwZywHKrE