daily-fix: dose-ladder checkpoint retention sizing rule
Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Overview / Motivation
Auto-filed by /daily 2026-07-07 (route 2) from the nightly transcript problem sweep.
Goal
add a between-rung checkpoint-retention clause to plan-compute-sizing.md: dose-to-band ladders keep only the dose-selected + latest checkpoints, cleaning non-selected rungs between rungs; plans size disk to the RETAINED set
Workflow gap
- Bug observed: #1112 (05000ba2): full-FT checkpoints ~15-28GB x 30 rungs filled the ft-7b default 200GB volume -> ENOSPC at rung 24/30, crash + terminate + fresh 800GB pod while billing $16/hr.
- Why it is a workflow gap: #1118 fixed volume-size threading (provisioning side); the plan-time retention rule is the missing complement that keeps footprints bounded regardless of volume size.
Proposed change
Add to .claude/rules/plan-compute-sizing.md (disk sizing): a dose-ladder / multi-rung training plan states its checkpoint-retention policy (keep dose-selected + latest only; clean non-selected rungs between rungs) and sizes disk to the RETAINED set; the critic REVISEs a ladder plan whose disk estimate assumes keeping every rung without justification. Upload-before-delete policy unchanged.
Scope / surfaces
- Primary target:
.claude/rules/plan-compute-sizing.md - Grep the workflow surface for the pattern before editing and update every hit.
Provenance
- Evidence: 05000ba2 (#1112) ENOSPC at rung 24/30; PM session e3f63b58 tracked the burn.