BLD-002 R6: a commit that changed two comment lines voided every release result (about 7 h) #4673

During the 0.70.1 release, the fix from 7d2dc28 to 04332f8 changed two comment lines in scripts/check_publish_preflight.sh, plus the fix's review record under evidence/. crates/, src/, Cargo.toml and Cargo.lock did not change. Every release result is tied to the commit, so all of them had to run again at the new commit: the rebuild and smoke on both GPUs, the dogfood, the clean-room, the probes and plants, and Lean T5. That is about 7 h for two comment lines.

Expected (BLD-002 R6): a result is tied to the inputs it read, not to the whole commit. A change that touches none of a row's inputs keeps that row's result. The receipt says the result was carried, from which commit, and the hash of the inputs that proves they are identical.

What the fix needs (no work until 0.70.1 is LIVE):

• Each long row declares its inputs: the files it reads, plus a build fingerprint (the tree hash of crates/, src/, Cargo.toml, Cargo.lock, the toolchain file and the tool versions).
• The receipt stores each row's input hash. A gate accepts a row from an older commit only when that hash is identical, and the row names the commit it came from.
• Rows whose inputs include a changed file run again (here: bashrs and the preflight self-test).
• A case table: a comment-only change to one script (only that script's rows run again), a change under crates/ (every row runs again), a toolchain change (every row runs again), an input the row did not declare (the row runs again).
• Fail closed: a row is carried only on an identical input hash. Anything not hashed runs again.
• This changes what "dogfood GO for this commit" means in the release gate, so the operator rules on it before it is enforced.

Reported by the operator (C295.5). Owner: build-kaizen.

Agent: aprender-27