Claude Code burns hours re-investigating a one-line numeric display-precision fix across sessions

A trivial defect — a single computed value displayed with one too many decimal places (7 vs 6, same underlying value) — consumed 2+ hours across multiple sessions without the fix landing.

Failure pattern:

1. No cross-session memory of the investigation. Each session re-derives the same diagnosis from scratch instead of resuming from the prior session's findings. The user pays for the same investigation repeatedly.

1. Over-investigation before a measurement. The agent reasoned in circles about the rounding rule for many tool calls instead of running the one computation that produces the actual number and ends the ambiguity in a single shot.

1. Built a fix theory on an unverified number. A prior session stated a figure; later sessions accepted it without re-checking, then reasoned on top of it.

1. Time/cost not proportional to defect size. A one-line display fix should not cost a long-running, multi-session budget.

Request: for small, bounded defects the agent should (a) take the one decisive measurement immediately rather than reasoning around it, (b) persist its diagnosis so a later session resumes instead of restarts, and (c) cap the effort it spends circling a one-line fix.