spec(nn): make specs/nn/hslm.t27 describe the trained HSLM (3 blocks, ctx 27, ReLU FFN with bias and alpha, embedding, vocab-729 head) #10062

What

specs/nn/hslm.t27 compiles and passes since #10055, but it does not describe the HSLM that gHashTag/trinity-training trained (src/hslm, commit 3e9648c1). Its own header lists the gaps:

• NUM_LAYERS 6; the trained model has DEFAULT_BLOCKS 3.
• CONTEXT_LEN 81; the best run (PPL 2.96, trinity docs/experiments/FOUND_EXPERIMENTS_SUMMARY.md) trained at context 27.
• FFN 243 -> 972 -> 243 with GELU, no bias, behind a second RMSNorm; the trained block is 243 -> 729 -> 243 with ReLU, a TWN alpha per matrix, a bias after each projection, and no norm.
• No token embedding (table + phi position encoding) and no vocab-729 output head (RMSNorm, ternary W_out, alpha, 1/243^(phi^-3), bias).
• The header says attention matches. It does not: the trained attention scales Q/K/V/O by TWN alphas and rotates Q and K with phi-RoPE.

Done when

• specs/nn/hslm.t27 describes the trained forward pass, cites the trained source files, and names what still differs.
• The backward pass is still checked by finite differences, including ReLU, the biases, the embedding and the head.
• specs/igla/own_model.t27 and its mirror specs/igla/own_model_board.t27 count it, and both pass.
• hslm, own_model and own_model_board report 0 FAIL and 0 vacuous on the Railway lab; HSLM and nn_HSLM are re-sealed.

This is the own-model board row "Match the spec to the trained HSLM".