diff --git a/.claude/skills/babysit-pipeline/SKILL.md b/.claude/skills/babysit-pipeline/SKILL.md
index f26aa105ef..f3021e206d 100644
--- a/.claude/skills/babysit-pipeline/SKILL.md
+++ b/.claude/skills/babysit-pipeline/SKILL.md
@@ -102,9 +102,13 @@ with evidence — never let the user ask "still running?".
 - **Defer, don't halt**: one stuck library must not stop a multi-spec
   queue — log it, cap ~90 min/spec, move on, and sweep leftovers at
   the end with targeted regens.
-- **Halt the whole queue** only on a failure CLUSTER (≥5 failed
-  `Generate:` runs in minutes = model daily quota exhausted; a
-  fallback model via `-f model=` can finish leftovers).
+- **Halt the whole queue** only on a failure CLUSTER: **≥3 distinct
+  (spec, library) pairs** failing within minutes of each other = model
+  daily quota exhausted; a fallback model via `-f model=` can finish
+  leftovers. The threshold is pairs, full stop — the raw run count
+  says nothing, because one impossible pair spends three runs on its
+  own auto-retries, so two bad pairs already produce six failures with
+  the pipeline entirely healthy (observed 2026-08-24).
 - Closing/merging pipeline PRs or issues yourself is out of bounds
   (CLAUDE.md external-write rule + mandatory workflow).
 
@@ -160,6 +164,24 @@ ggplot2/network-force-directed, all three of which had been read as
 capability gaps. Two failures on the same pair → `deferred.log` with
 the reason, move on, sweep at the end.
 
+Check the run list before deferring, though: since #10627 the
+workflow spends its own three attempts per campaign, so a pair that
+comes back missing may already have failed three times, and the
+manual retry adds nothing. There is no per-spec filter on the API, so
+filter the output by run title:
+
+```bash
+gh run list --workflow=impl-generate.yml --limit 40 \
+  --json conclusion,createdAt,displayTitle \
+  --jq '.[] | select(.displayTitle | test("<spec>")) | "\(.createdAt) \(.conclusion) \(.displayTitle)"'
+```
+
+Three `Generate: <lib> for <spec>` failures minutes apart is a measured
+gap; one failure is a flake worth the retry. The same check covers
+`RESULT=TIMEOUT` with `recent generate failures: 0`, which only means
+the failures fell outside the driver's 25-minute window, not that
+nothing failed.
+
 ## Gotchas
 
 - **`Marking <lib> as failed: N generation attempts` counts more than
@@ -185,12 +207,25 @@ the reason, move on, sweep at the end.
   failed 17:55, retried and succeeded 18:02, no human involved), so a
   single occurrence is noise — only a repeat on the same pair means
   anything.
-- **Static library + interactive/3D spec is the one gap that is
-  usually real.** plotnine, pygal, seaborn, matplotlib and ggplot2
-  against `*-interactive`, `*-drilldown`, `*-realtime`, `*-streaming`,
-  `slider-*`, `*-3d` specs make up most genuine dead ends; 18 of the
-  45 real gaps behind those labels were exactly this shape. Still
-  give them the one retry, then defer without further ceremony.
+- **Do not predict capability gaps — measure them.** The 2026-08-24
+  backfill guessed four times which pairs were genuinely impossible
+  (chartjs on treemap/sankey, plotnine on 3D, ggplot2 on wireframe,
+  "static library vs. interactive spec") and every category-level
+  guess was wrong: 17 of 20 parked pairs generated fine, pygal and
+  chartjs each succeeded on the very plot types they were written off
+  for, and `bar-3d-categorical` succeeded in plotnine while
+  `scatter-3d` did not. Only three pairs failed with a full budget —
+  plotnine on `scatter-3d`, `contour-3d`, `line-3d-trajectory`, all
+  needing a spatial projection plotnine does not have, while the
+  "3D" spec that is representable in 2D went through. A gap is real
+  when three attempts under a fresh campaign budget say so, never
+  because the pairing sounds implausible.
+- **A spec id from an issue title may not exist.** `impl:*:failed`
+  labels outlive their specs: of 26 spec IDs harvested that way, 14
+  had no `plots/<spec>/` on main at all, and a dispatch for one dies
+  at `Validate specification exists` seconds in. Intersect any
+  label-derived list with `git ls-tree -d origin/main plots/` before
+  queueing it.
 - **The scripts' library list is a copy** of `core/constants.py`'s
   registry (15 libs). When a library is added/removed, update
   `monitor_spec.sh`'s `ALL_LIBS` and `run_spec.sh`'s `lang_of` in the
diff --git a/.claude/skills/babysit-pipeline/run_spec.sh b/.claude/skills/babysit-pipeline/run_spec.sh
index 7a8c8cec8f..3591e310cc 100755
--- a/.claude/skills/babysit-pipeline/run_spec.sh
+++ b/.claude/skills/babysit-pipeline/run_spec.sh
@@ -77,7 +77,24 @@ recent_generate_failures() {
     || echo "?"
 }
 
-git -C "$REPO" fetch origin main --quiet
+# Two drivers polling the same checkout collide on `.git/refs/remotes/origin/main`
+# ("cannot lock ref ... is at X but expected Y"). A lost fetch leaves origin/main
+# stale, which makes meta_present understate progress — so retry, and say so in
+# the log when all three attempts lose, rather than silently polling a stale ref.
+fetch_main() {
+  local i err
+  for i in 1 2 3; do
+    # Keep stderr: a ref-lock collision and, say, an auth or network failure
+    # need different responses, and discarding the message hides which it was.
+    err=$(git -C "$REPO" fetch origin main --quiet 2>&1) && return 0
+    sleep $(( i * 3 ))
+  done
+  echo "[warn] git fetch origin main failed 3x — origin/main may be stale this round" | tee -a "$LOG"
+  echo "[warn]   last error: ${err}" | tee -a "$LOG"
+  return 1
+}
+
+fetch_main
 TODO=()
 for lib in "${LIBS[@]}"; do
   if meta_present "$lib"; then
@@ -106,7 +123,7 @@ iters=$(( MAXMIN * 60 / INTERVAL ))
 idle=0; last_done=-1
 for ((i=1; i<=iters; i++)); do
   sleep "$INTERVAL"
-  git -C "$REPO" fetch origin main --quiet
+  fetch_main
   missing=(); done_n=0
   for lib in "${TODO[@]}"; do
     if meta_present "$lib"; then done_n=$((done_n+1)); else missing+=("$lib"); fi
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 8b7f01943a..b6ec19a528 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -180,6 +180,21 @@ aggregate instead: an italic *Catalog* line at the end of the version section an
 
 ### Changed
 
+- **`babysit-pipeline`'s gap-backfill guidance corrected against a completed run** — the
+  section shipped in #10628 was written mid-backfill and three of its claims did not
+  survive the rest of it. The halt-on-cluster threshold counted raw failed runs, but one
+  impossible pair now burns three runs on its own retries, so two bad pairs tripped it
+  while the pipeline was healthy; it counts distinct `(spec, library)` pairs now. The
+  "static library against an interactive or 3D spec is the gap that is usually real"
+  gotcha is replaced by its opposite: every category-level prediction made during the
+  backfill was wrong — 17 of 20 parked pairs generated fine, and `bar-3d-categorical`
+  succeeded in plotnine while `scatter-3d` did not — so a gap counts as real only after
+  three attempts under a fresh budget. Added: spec IDs harvested from issue titles must be
+  intersected with the actual `plots/` directories (14 of 26 pointed at specs that no
+  longer exist), and a `TIMEOUT` with zero recent failures means the failures aged out of
+  the driver's window, not that nothing failed. `run_spec.sh` also retries its
+  `git fetch`: two drivers polling one checkout collide on the `origin/main` ref lock,
+  and a lost fetch made the poller read stale metadata and understate progress (#10660).
 - **`babysit-pipeline` learned gap backfill** — closing coverage gaps across the catalogue is
   a different job from watching one fresh spec, and the skill only described the latter. It
   now documents reading the missing set from `plots/*/metadata/` instead of from labels