launcher: adopt D-018 -- no gradient checkpointing, push every 127 steps, plan at the sustained 12,300 tok/s

launcher: repin the E-044 trainer

launcher: repin the instrumented trainer

launcher: repin train files; print TRAIN_ARGV under PREP_ONLY so the rehearsal can parse the real flag list

launcher: repin hubckpt to the bounded read-back

phase4_session: repin the trainer

phase4_session: repin the trainer (rank-visible eval failure, cross-rank peak memory)

E-037: torchrun's first positional IS the script -- passing sys.executable made it compile the python binary and die with 'source code cannot contain null bytes' in 8.6s. Affects the Phase 4 launcher identically at session 1. Probe also split into smoke (20 steps) and soak (180) modes

train: escape the literal percent signs in two --help strings (argparse interpolates help text, so '+52 %' crashed --help with 'must be real number, not dict' -- caught by the CPU rehearsal's stage 3)

perm_sha: identify the visit order by hashing the permutation, not just its seed (E-035 residual) -- cursor gains shuffle_perm_sha, asserted on resume; session repins the two files

phase4_session: plan() tiles to the horizon, kill the process group, df_free_gb helper, plan before the pull, session-relative TRAIN timeout

phase4_session: default to P3's measured planning rate (9358 tok/s) and the 6.9 h session the launch plan now books

mix_fingerprint: fold the manifest's per-shard sha256 into the corpus identity (names+sizes could collide across rebuilds); phase4_session repins shard_dataset

pin: kernels/phase4_session.py (RepoMissing resume branch, disk gate, PHASE4_PREP_ONLY)

Upload kernels/phase4_session.py with huggingface_hub

Upload kernels/phase4_session.py with huggingface_hub