feat(ride): displacement arm - ridge-projected teacher+base hidden regression (run-016)

[ronaldtse]

The RIDE displacement arm (TODO.sota-2026/05 in this repo), built and launched after the direction-transferability probe passed decisively (max-layer cos 0.9513, domain-general L0-L8).

Mechanism: regress the student's encoder hidden states toward ridge-projected displaced teacher targets h_t + lam*(h_t - h_b), with the frozen r7 teacher and r6 base both in the training loop. The per-layer ridge map is fit at arm start from the teacher's hidden space to the student's INITIAL hidden space on the same inputs - by linearity the displacement survives the projection. This is the first student-side lever with a mechanism novel to the ledger (representation-space extrapolation; all eight prior levers were loss/data/optimizer-side) and a measured transfer premise, per the closure rule in TODO 03.

• src/gpu/ride.py - displaced_targets, masked_mse, closed-form fit_ridge (TDD, 12 tests, tests/test_ride.py)
• modal_distill.py - ride_aux spec wiring: frozen base alongside teacher (byte-table parity asserted at startup), ridge prefit on the first fit_batches batches, sub-batched frozen passes (memory), beta auto-calibrated so the MSE term is 10% of CE at start (pre-registered), ride.pt persisted with checkpoints for resume
• distill_specs.yaml - ara-diac-small-2-1-ride (run-016): the 2.1 recipe verbatim (r7 labels, Muon, 6 epochs, seed 42) + ride_aux layers 0-8, lam 1.0. Gate: adopt iff >= 0.3pp better than 4.5701 (E4-style bar). In flight.