v1.2 — comprehensive audit and convention parity

Deep Learning: Making It Learnable v1.2

Version 1.2 is the comprehensive audit and convention-parity release.

• The HTML edition remains canonical; the attached PDF is its fixed 552-page print
  conversion.
• All three interludes now have independent figure namespaces, unnumbered equations,
  closing retrieval checks, and the same public authoring conventions as numbered
  chapters.
• The attention-as-test-time-regression interlude makes three memory solvers
  inspectable without overstating their relation to softmax attention or state-space
  models.
• Book-wide Plan → Code, exercise-tag, source, book-voice, frozen-output, and PDF
  text-layer contracts are enforced in CI.
• The exercise bank and evidence apparatus expand across optimization,
  generalization, sequence modeling, attention, pretraining, calibration, and
  generative modeling.
• The temperature thread now runs from kernel bandwidth, through learned similarity
  scale, to CLIP's training-time logit scale and a frozen model's post-hoc calibration.
• The epilogue owns the E. figure namespace, cites its test-time-control and
  mixture-of-experts sources directly, and leaves its control equations unnumbered.

At release, all 133 retained stdout blocks are byte-identical across HTML and TeX,
all manuscript and Python audits pass, no missing glyphs or corrupted PDF text are
detected, and every internal cross-reference resolves.

PDF SHA-256: 264263182a9d601460ed5f061aad62ee6d6933e1566fea8c26d9685dc8e4d10b