undefined

Persist `PRIOR_FIX_RANGE` (and `$LAST_FIX_SHA` as the next iter's `$PREV_HEAD`) into the babysit state file or session env. If the iter pushed a single commit, `FIRST_FIX_SHA == LAST_FIX_SHA` and the range collapses to `<sha>^..<sha>`.

If this iter pushed zero commits (CI re-run only) → no fix range to record; skip the Gate B self-introduced check for the next iter, but still run Gate A as normal.

**Why not compute lazily at Gate B**: computing at push time anchors the range to the exact commits that addressed iter (N-1) findings. Lazy computation at Gate B time could pick up unrelated commits if the user manually edits the branch between iters.

gh api repos/$OWNER/$REPO/pulls/$N/comments \
  --jq '[.[] | select(.body | contains("<!-- pr-review:finding-id=")) |
         {id, created_at, path, line, body,
          justification: (.body | capture("<!-- pr-review:justification=(?<j>[^ ]+) -->").j),
          race_meta: (.body | capture("\\[window=(?<w>[^,]+), damage=(?<d>[^,]+), recovery=(?<r>[^\\]]+)\\]") // null)}]'

Status: needs-user-input (diminishing returns)

This iter's pr-review surfaced only hygiene findings — no Reachable / Precedent /
Asymmetric / Historical justification on any new finding.

Hygiene followups (N):
  <list — id, slug, file:line, one-line failure mode>

Continuing the loop will likely surface more hygiene from the same code paths.

Your call:
  (s) ship — open a single follow-up issue collecting the hygiene items, mark PR ready-to-merge
  (p) polish — keep looping (override the gate for this round)
  (r) re-review-full — challenge whether the self-loop missed anything (force `mode=full` on next pr-review)

git diff --name-only $prior_fix_range                  # files touched
git diff -U0 $prior_fix_range -- <file>                # line-level attribution

Status: convergence-audit (race-of-race detected)

iter (N-1) fix surface attracted N race-class findings this iter (cluster):
  <id> <slug> @ <file:line>  window=<w> damage=<d> recovery=<r>
  ...

Audit verdict per finding:
  <id>: modify    — <reason: Asymmetric / mitigation safe / etc>
  <id>: wontfix   — <five-field summary from Wontfix Template>
  <id>: defer     — <followup issue suggestion>

Your call:
  (a) accept all verdicts (post wontfix replies via template, address modify items, open defer issues)
  (m) modify a specific verdict — say which finding-id and target verdict
  (s) ship — accept all wontfix + defer as-is, mark PR ready-to-merge
  (p) override audit — treat as normal iter, loop back to step 1

Wontfix — deliberate trade-off.

Race window: <ms / s / min / hr> between <op A> and <op B>.
Precondition: <only fires when X is in Y state for N+ time>
Damage if race fires: <not data-loss / not deadlock / only X happens N seconds earlier than ideal>
Recovery path: <new event / cron sweeper / next webhook covers it; user-visible behavior unchanged>

Asymmetric check: <not security / not data-loss / not data-integrity / not billing>
Mitigation cost: <atomic re-check / two-step merge into transaction is doable, but introduces new race-of-race surface at X>

Acknowledged as known trade-off; fault tolerance covers genuinely <abandoned / stranded / dropped> class.
Tracking: <if needed, opened followup issue X>

PR/MR: <link>
Status: ready-to-merge | needs-user-input | blocked
Checks: <green>/<total>
Addressed (this run): <list of SHA → comment ref + one-liner>

Awaiting your decision:
  Discuss (I did NOT reply): <list with comment text + my read of the ambiguity>
  Out-of-scope: <list>  → open follow-up issues for any of these? (y/N per item)

Blockers (if any): <description + what I tried>

Next command: gh pr merge --squash <id>   # or: glab mr merge <id>