[cskwork]

The four-phase flow now uses safe disposable probes for empirical uncertainty, requires evidence matched to performance and behavior-preservation claims, and tests a shared premise before a third fix assumes it again. Plan review checks the proposed proof method; Build and QA check the results.

The guidance lives in five existing instruction files. Redundant prose was shortened so the standard prompt corpus decreases from 5,998 to 5,993 words and every phase bundle remains below 2,500 words. The entrypoint, scripts, approval gates, retry caps, schemas, templates, and version are unchanged. Sustained optimization remains opt-in with an agreed target and effort limit.

Validation:

• bash tests/selftest.sh: PASS, 127 checks.
• git diff --check: clean.
• Independent review: PASS after exercising six instruction scenarios, including noisy measurements, compile-only refactor evidence, and unrelated failed fixes.

Adapted selected practices from Poteto mode to the existing SDLC flow.