v0.16.0

BREAKING CHANGE

• a subprocess worker that dies mid-suite is now replaced and the
  run continues, where previously every remaining MLP failed. Suite scores for
  submissions that crash on some MLPs will improve, because only the crashing MLPs
  are now scored as failures.
• the no-dataset default contest is now 1024x16 with 200,000
  ground-truth samples and an 8 GB solution-process memory limit, replacing 256x32
  with 2,560,000 samples and 64 GB. Local scores from before this change are not
  comparable, and an estimator hardcoding 256x32 will now see 1024x16.
• the default flop_budget moves from 272_000_000_000 to
  2**41 (2,199,023,255,552). Runs that relied on the old default without passing
  --flop-budget will meter against a ~8x larger cap.
• ContestSpec.lambda_flops_per_second now defaults to 0.0 instead
  of 1e11, so effective_compute() and is_combined_budget_exhausted() return
  FLOP-only results unless a rate is passed. Callers re-scoring an earlier round
  must pass PHASE1_LAMBDA_FLOPS_PER_SECOND explicitly and set
  residual_wall_time_limit_s=None.

Feat

• budget: keep every round's config side by side, and document the progression
• seeds: add seed-protocol 4.0 with keyed per-MLP derivation
• config: default to the graded 1024x16 shape and the 8 GB solution memory cap
• budget: default the per-MLP budget to 2**41
• budget: default lambda to 0 so residual wall time is gated, not priced
• limits: set the predict wall cap to 120 s and make both caps configurable
• starterkit: detect dependency drift against the starter kit
• dataset_torch: record the bake's chunking in the synthesized breakdown

Fix

• docs: declare rounds and gpu-dataset-generation in the llms.txt manifest
• docs: add frontmatter to the rounds mirror so the site builds
• scoring: let make_contest_from_dataset take an explicit seed_salt
• scoring: bound the metered window without billing harness checks
• runner: restart a dead worker so one death does not fail the whole suite
• docs: Phase 1 re-score recipe must also pin the 60 s wall cap
• cli: report packaging failures with a packaging code, not a scoring one
• scoring: stop charging harness transport to the per-MLP wall clock
• packaging: scope the secret-file preview to submission candidates
• packaging: name dropped siblings, refuse a single file that imports one
• limits: report wall-clock exhaustion correctly and cap residual at the graded 400 ms
• domain: pin MLP weights to float32 on every construction path
• submit: ask AIcrowd whether a submission is allowed
• dataset_torch: emit operations, use the real chunk_size, drop the fake budget
• card: an empty companion_repo must omit the cross-link, not invent one