diff --git a/.github/scripts/fhir-bench/search_by_shape.py b/.github/scripts/fhir-bench/search_by_shape.py
new file mode 100755
index 000000000..5142672a1
--- /dev/null
+++ b/.github/scripts/fhir-bench/search_by_shape.py
@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+#
+# Per-query-shape search latency attribution — replaces a `jq -rs` program
+# that slurped the whole k6 `--out json` point stream into memory in one go.
+#
+# Called from: the `benchmark` job's "Attribute search latency per query
+# shape" step (fhir-benchmark.yml), via:
+#   python3 "$GITHUB_WORKSPACE/.github/scripts/fhir-bench/search_by_shape.py" \
+#     "$POINTS" > "$OUT"
+#
+# Why this streams instead of slurping (`jq -rs` reads the entire NDJSON
+# stream into one in-memory array before doing anything else): the
+# self-hosted runner has been lost during this exact step three times —
+# runs 30550776427, 33515369645 and 36423973848. The last of those was
+# backend=sqlite, tests=prewarm,search: searching a near-empty database is
+# very fast, so k6 emits an enormous number of points in a short wall-clock
+# window and the points file balloons — that run's artifact was 317 MB
+# compressed, and run 36412022609's 854 MB points file only just made it
+# through before `jq -rs` started failing this way on bigger ones. This
+# script reads the file one line at a time and keeps only the (few) floats
+# it needs per query shape, so memory stays proportional to the number of
+# distinct shapes (~21), not to the number of requests in the run.
+#
+# Input: sys.argv[1] — a k6 `--out json` NDJSON file, one JSON object per
+# line, every metric of every request (Point/http_req_duration is one of
+# many). A line that isn't valid JSON (a runner killed mid-write, e.g.) is
+# skipped rather than aborting the whole file, unlike `jq -s`, which fails
+# the entire slurp on the first parse error.
+# Output: stdout — a Markdown table, one row per query shape
+# (`<searchType> <resourceType>?<name>`), sorted by p95 descending, with
+# n/median/p95/max in milliseconds. With no matching points, only the two
+# header lines are printed (matching what the old jq program produced for
+# an empty group list).
+#
+# Reproduces the jq program's semantics exactly, including two quirks that
+# would silently diverge under naive Python equivalents:
+#   - jq's `//` treats null AND false as "missing" (not just null) when
+#     falling back to "?" for an absent tag; an empty string is a value and
+#     is NOT replaced. tag_or() below implements that, not `dict.get(k, "?")`.
+#   - jq's `round` rounds half away from zero (round(2.5) == 3,
+#     round(-2.5) == -3); Python's built-in round() rounds half to even
+#     (round(2.5) == 2), which would silently misreport rounded milliseconds
+#     ending in .5. jq_round() below reimplements the C `round()` behaviour.
+# median/p95/max use the same indices as the jq program: med = sorted[n//2],
+# p95 = sorted[min(floor(n*0.95), n-1)], max = sorted[n-1]. Ties on p95 sort
+# by ascending shape name, matching jq's `group_by(.shape)` (ascending)
+# followed by a stable `sort_by(-.p95)`.
+import json
+import math
+import sys
+from array import array
+
+
+def tag_or(tags, key):
+    """jq's `(.data.tags.KEY // "?")`: null and false both count as
+    missing; an empty string does not."""
+    val = tags.get(key)
+    if val is None or val is False:
+        return "?"
+    return val if isinstance(val, str) else str(val)
+
+
+def jq_round(x):
+    """jq's `round`: round half away from zero, unlike Python's
+    round-half-to-even."""
+    if x >= 0:
+        return math.floor(x + 0.5)
+    return math.ceil(x - 0.5)
+
+
+def main():
+    if len(sys.argv) != 2:
+        print("usage: search_by_shape.py <points.json>", file=sys.stderr)
+        return 2
+
+    path = sys.argv[1]
+    by_shape = {}
+
+    try:
+        with open(path, "r", encoding="utf-8", errors="replace") as fh:
+            for line in fh:
+                # Cheap pre-filter before paying for json.loads: every
+                # http_req_duration Point line contains both substrings.
+                if '"Point"' not in line or '"http_req_duration"' not in line:
+                    continue
+                try:
+                    obj = json.loads(line)
+                except ValueError:
+                    continue
+                if not isinstance(obj, dict):
+                    continue
+                if obj.get("type") != "Point" or obj.get("metric") != "http_req_duration":
+                    continue
+                data = obj.get("data") or {}
+                value = data.get("value")
+                if not isinstance(value, (int, float)) or isinstance(value, bool):
+                    continue
+                tags = data.get("tags") or {}
+                shape = (
+                    tag_or(tags, "searchType") + " " +
+                    tag_or(tags, "resourceType") + "?" +
+                    tag_or(tags, "name")
+                )
+                values = by_shape.get(shape)
+                if values is None:
+                    values = array("d")
+                    by_shape[shape] = values
+                values.append(float(value))
+    except OSError as exc:
+        print(f"search_by_shape.py: {exc}", file=sys.stderr)
+        return 1
+
+    print("| Query shape | n | median (ms) | p95 (ms) | max (ms) |")
+    print("|---|---:|---:|---:|---:|")
+
+    # Ascending shape order first (mirrors jq's group_by(.shape)); the
+    # later sort by -p95 is stable, so ties keep this ascending order,
+    # matching jq's group_by + stable sort_by(-.p95).
+    rows = []
+    for shape in sorted(by_shape.keys()):
+        s = sorted(by_shape[shape])
+        n = len(s)
+        med = s[n // 2]
+        p95 = s[min(math.floor(n * 0.95), n - 1)]
+        mx = s[n - 1]
+        rows.append((shape, n, med, p95, mx))
+
+    rows.sort(key=lambda r: -r[3])
+
+    for shape, n, med, p95, mx in rows:
+        print(f"| `{shape}` | {n} | {jq_round(med)} | {jq_round(p95)} | {jq_round(mx)} |")
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())
diff --git a/.github/workflows/fhir-benchmark.yml b/.github/workflows/fhir-benchmark.yml
index f3a644267..81d868627 100644
--- a/.github/workflows/fhir-benchmark.yml
+++ b/.github/workflows/fhir-benchmark.yml
@@ -1736,49 +1736,37 @@ jobs:
       # Without this the search suite reports a single aggregate latency and we
       # cannot tell which of the ~21 query shapes is slow. Produces a table sorted
       # by p95 so the worst shape is the first line.
+      #
+      # This used to be a `jq -rs` program, which slurps the whole NDJSON point
+      # stream into memory before doing anything else. The self-hosted runner
+      # was lost during this exact step three times — runs 30550776427,
+      # 33515369645 and 36423973848 (the last: backend=sqlite,
+      # tests=prewarm,search — searching a near-empty database is very fast, so
+      # the points file was several GB; the artifact was 317 MB compressed, and
+      # run 36412022609's 854 MB file only just got through). search_by_shape.py
+      # reads the file line by line instead and reproduces the same jq
+      # semantics (see its header) — timeout-minutes below is a backstop, not
+      # the fix.
       - name: Attribute search latency per query shape
         if: always()
+        timeout-minutes: 15
+        env:
+          BACKEND: ${{ matrix.backend }}
         run: |
           set -euo pipefail
-          POINTS="$GITHUB_WORKSPACE/bench-results/${{ matrix.backend }}/search-points.json"
+          POINTS="$GITHUB_WORKSPACE/bench-results/$BACKEND/search-points.json"
           [ -f "$POINTS" ] || { echo "No search points file — search suite did not run."; exit 0; }
 
-          OUT="$GITHUB_WORKSPACE/bench-results/${{ matrix.backend }}/search-by-shape.md"
-
-          # k6's `--out json` emits one JSON object per line. Keep the
-          # http_req_duration points, group by the tags the suite already sets, and
-          # compute count/median/p95/max per shape.
-          jq -rs '
-            [ .[]
-              | select(.type == "Point" and .metric == "http_req_duration")
-              | { shape: ((.data.tags.searchType   // "?") + " " +
-                          (.data.tags.resourceType // "?") + "?" +
-                          (.data.tags.name         // "?")),
-                  v: .data.value } ]
-            | group_by(.shape)
-            | map(
-                (sort_by(.v))    as $s
-                | ($s | length)  as $n
-                | {
-                    shape: $s[0].shape,
-                    n:     $n,
-                    med:   $s[($n / 2 | floor)].v,
-                    # Clamp the percentile index to the last element; with only a
-                    # handful of samples in a shape, floor(n*0.95) can land past the
-                    # end. Bind $n explicitly rather than calling `length` inside the
-                    # index expression — there `.` is the index number, not the
-                    # array, and `length` would silently return the number itself.
-                    p95:   $s[([($n * 0.95 | floor), ($n - 1)] | min)].v,
-                    max:   $s[$n - 1].v
-                  }
-              )
-            | sort_by(-.p95)
-            | "| Query shape | n | median (ms) | p95 (ms) | max (ms) |",
-              "|---|---:|---:|---:|---:|",
-              (.[] | "| `\(.shape)` | \(.n) | \(.med | round) | \(.p95 | round) | \(.max | round) |")
-          ' "$POINTS" > "$OUT" || { echo "::warning::per-shape attribution failed"; exit 0; }
-
-          echo "### Search latency by query shape (${{ matrix.backend }})" >> "$GITHUB_STEP_SUMMARY"
+          OUT="$GITHUB_WORKSPACE/bench-results/$BACKEND/search-by-shape.md"
+
+          # Keep the http_req_duration points, group by the tags the suite
+          # already sets, and compute count/median/p95/max per shape — see
+          # search_by_shape.py's header for why this streams instead of
+          # slurping, and for the jq-equivalence quirks (the `//` "missing"
+          # semantics and round-half-away-from-zero) it reproduces exactly.
+          python3 "$GITHUB_WORKSPACE/.github/scripts/fhir-bench/search_by_shape.py" "$POINTS" > "$OUT" || { echo "::warning::per-shape attribution failed"; exit 0; }
+
+          echo "### Search latency by query shape ($BACKEND)" >> "$GITHUB_STEP_SUMMARY"
           cat "$OUT" >> "$GITHUB_STEP_SUMMARY"
           cat "$OUT"