Phase 2 — Uniform + cued arms with verified equal replay budget

[Tejas-JB]

PRD §6 Phase 2. Blocked by Phase 1 gate.

Deliverables

• Replay-pass implementation with configurable budget B (measured in training tokens, not example count — see docs/decisions.md) and flag-oversampling ratio r (default 5×).
• Hard invariant: assert at every replay step that
  |tokens(uniform_batch) − tokens(cued_batch)| / tokens(uniform_batch) < 0.01.
  Log every batch's token count. Without this assertion the "equal compute" claim doesn't hold.
• Uniform arm, ≥3 seeds.
• Cued arm, ≥3 seeds (same seed list as uniform, for paired comparison).
• One Modal job at a time (owner-confirmed sequential parallelism — integrity > speedup).

Exit criterion

All three arms (no-replay from Phase 1, uniform, cued) have complete retention curves across ≥3 seeds with verified-equal replay budgets between uniform and cued.