The review gate re-ran almost every selected witness on every round even
when the executor's fix was twelve lines: the ledger fingerprint and the
diff selection both worked on whole files, and the card hosts are
thirteen thousand lines each. On #500 those twelve lines in
houseplan-editor-runtime.ts pulled 53 of the 75 witnesses the third
round ran, and the gate cost 140 job-minutes and an hour of the
reviewer's wall clock across three rounds.
The patch side is now judged by the anchor's neighbourhood — the anchor
lines plus ANCHOR_RADIUS_LINES on each side — in both the ledger
fingerprint and the diff selection, which now reads hunk ranges from
git diff --unified=0. The guard side keeps whole-file granularity: a
guard has no anchor and changes as a whole. An anchor that is not found
exactly once falls back to the whole file, and so does a file whose
hunks were not read: not knowing is not proof. Same class of
approximation as the existing diff selection, with the nightly full
gate (#513) as the floor.
Two more cuts to the wall clock of a review round. The shard plan is now
computed before the environment is installed — restore the ledger,
select, split, and only then pay for npm ci, Python and Chromium; the
job still runs, so the review gate's proof (#510) is unaffected. And the
matrix goes from three shards to six: the same job-minutes, half the
wall time.
On the #500 round the selection drops 60 → 7. Four witnesses guard the
new logic, including the two unsafe defaults (ambiguous anchor, missing
hunks).
Issue: #518
User-Visible: no
Codex
2783ceff94