Re-read the central files immediately before writing or reviewing a plan; do
not rely on earlier conversation reads when code may have changed.
For renames, refactors, and callsite sweeps, derive the inventory from
tracked files with git ls-files plus rg. Be explicit about exclusions and
rerun the same search as verification.
Before planning recovery or cleanup recipes, verify every step against the
current cmd_* / plan_* code and the relevant tool or kernel behavior.
Treat issue recipes as hypotheses until the code proves them.
Architecture docs describe behavioral contracts, not internal helper names.
Verify wrapper process/lifetime claims from the wrapper code before writing
docs that depend on them.
For external-tool exit-code or wording classifiers, trace the specific
subcommand return path in reference/; a shared errno table is not enough to
prove one invocation’s behavior.