fix(agent): read the drawn texts of every library for the reply

.github/workflows/ci-image.yml

@@ -586,15 +586,15 @@ jobs:
           echo "$CHANGED_FILES"
 
           # What the image is built from: its Dockerfile, the service package,
-          # the harness it copies, the dependency lock, README.md (copied next
+          # the harness and the R and Julia text probes it copies, the dependency lock, README.md (copied next
           # to pyproject.toml), the catalogue's Julia project (the toolchain's
           # Julia stage precompiles it), the catalogue's JavaScript harness and its
           # lock (the toolchain's Node stage installs them) and the root .dockerignore
           # (the build context); and the agents package's copy of the import allowlist
           # and the JavaScript error grammar, which a step compares with the image, and
           # the export page's wrapper and the API's preview relay, which steps probe in the
           # image's Chromium.
-          RENDERER_CHANGES=$(echo "$CHANGED_FILES" | grep -E '^(agents/renderer/|agents/__init__\.py$|agents/anyplot/render/harness\.py$|agents/anyplot/render/convert\.py$|agents/anyplot/render/page\.py$|api/preview_relay\.py$|agents/anyplot/code/profiles/javascript/(import-allowlist\.json|errors\.py)$|tests/image/|tests/unit/agents/runtime/convert_fixtures\.py$|pyproject\.toml$|uv\.lock$|README\.md$|Project\.toml$|Manifest\.toml$|package\.json$|package-lock\.json$|automation/js-render/|\.dockerignore$|\.github/workflows/ci-image\.yml$)' || true)
+          RENDERER_CHANGES=$(echo "$CHANGED_FILES" | grep -E '^(agents/renderer/|agents/__init__\.py$|agents/anyplot/render/harness\.py$|agents/anyplot/render/convert\.py$|agents/anyplot/render/text_probe\.(R|jl)$|agents/anyplot/render/page\.py$|api/preview_relay\.py$|agents/anyplot/code/profiles/javascript/(import-allowlist\.json|errors\.py)$|tests/image/|tests/unit/agents/runtime/convert_fixtures\.py$|pyproject\.toml$|uv\.lock$|README\.md$|Project\.toml$|Manifest\.toml$|package\.json$|package-lock\.json$|automation/js-render/|\.dockerignore$|\.github/workflows/ci-image\.yml$)' || true)
 
           if [[ "$EVENT_NAME" == "workflow_dispatch" && "$FORCE_RUN" == "true" ]]; then
             echo "Manual trigger with force_run=true, will build the renderer image"
@@ -1067,6 +1067,12 @@ jobs:
           width, height = struct.unpack(">II", png[16:24])
           assert (width, height) == (3200, 1800), (width, height)
           print(f"OK: R plot-light.png {width}x{height}, {len(png)} bytes")
+          # The row's text probe (R_TEXT_PROBE) left the drawn texts, which the harness made the probe.
+          import json
+          labels = {(label["kind"], label["text"]) for label in json.load(open("work-r/probe-light.json"))["labels"]}
+          assert ("title", "Größe – Überblick") in labels, labels
+          assert ("x tick label", "b") in labels and ("x axis label", "x") in labels, labels
+          print(f"OK: R probe with {len(labels)} drawn texts")
           EOF
           tail -n 1 work-r/harness.out
 
@@ -1084,12 +1090,12 @@ jobs:
           set -euo pipefail
           mkdir -p work-jl
           python3 - > work-jl/row.flags <<'EOF'
-          from agents.renderer.commands import JULIA_WARMUP, command_for
+          from agents.renderer.commands import JULIA_TEXT_PROBE, JULIA_WARMUP, command_for
 
           row = command_for("julia", "makie")
           assert row.exec_deny and not row.rlimit_as, "the Julia row runs under the exec filter, without RLIMIT_AS"
           assert "--compiled-modules=strict" in row.argv, row.argv
-          assert row.argv[-2:] == ("-L", JULIA_WARMUP), row.argv
+          assert row.argv[-4:] == ("-L", JULIA_WARMUP, "-L", JULIA_TEXT_PROBE), row.argv
           for key, value in row.env.items():
               print(f"-e\n{key}={value}")
           EOF
@@ -1126,6 +1132,11 @@ jobs:
           width, height = struct.unpack(">II", png[16:24])
           assert (width, height) == (3200, 1800), (width, height)
           print(f"OK: Makie plot-light.png {width}x{height}, {len(png)} bytes")
+          # The row's text probe (JULIA_TEXT_PROBE) left the figure's drawn texts, which the harness made the probe.
+          import json
+          kinds = {label["kind"] for label in json.load(open("work-jl/probe-light.json"))["labels"]}
+          assert {"x tick label", "y tick label"} <= kinds, kinds
+          print(f"OK: Makie probe with the kinds {sorted(kinds)}")
           EOF
           tail -n 1 work-jl/harness.out
 

agents/anyplot/grounding.py

@@ -7,12 +7,17 @@
 
 Two measurements, both code:
 
-* `drawn_texts` reads the drawn texts the render's probe reports (`harness.drawn_labels`,
-  the `labels` field) by role, every one of them; `plot_texts` clips them into
-  `PlotShows.texts`, which the root reads fenced next to the result. The probe is written by
-  the code under test, so every value is checked for type and clipped; a probe without
-  `labels` (a renderer older than the field, or a failed text report) yields None, and a
-  run with requested texts then carries `UNVERIFIED_LINE` instead of a silent pass.
+* `drawn_texts` reads the drawn texts the render's probe reports (the `labels` field) by
+  role, every one of them; `plot_texts` clips them into `PlotShows.texts`, which the root
+  reads fenced next to the result. Every library's render reports them since 2026-11-01:
+  matplotlib, seaborn and plotnine through `harness.drawn_labels`, pygal, altair and
+  lets-plot through `harness.install_text_probes`, the JavaScript libraries, plotly and bokeh
+  through the browser harness (`render.mjs`, `writeTextProbe`), ggplot2 and Makie through the
+  text probes their rows load (`text_probe.R`, `text_probe.jl`, `harness.collect_text_probe`).
+  The probe is written by the code under test, so every value is checked for type and
+  clipped; a probe without `labels` (a failed text report) yields None, which the root never
+  sees (`plugins/tool_safety.fence_pipeline_notes` drops it), and a run with requested texts
+  then carries `UNVERIFIED_LINE` instead of a silent pass.
 * `expected_misses` checks each of the run's `expected_texts` (the texts and number formats
   the user asked to see, verbatim) against all drawn texts, never the clipped summary. A
   text (`Population 2025`, `Top 10`) must stand whole in one drawn text (case, spacing,
@@ -23,12 +28,13 @@
   Each miss becomes a residual line, so the result says it honestly, and with
   `AGENT_EXPECT_CHECK=repair` a measured finding the repair round gets.
 
-A render without drawn texts (R, altair, pygal and lets-plot, which write no probe) is checked by the
-reviewer instead where its runtime says so (`RuntimeAdapter.review_texts`, owner decision of
-2026-10-29): the review request carries the requested texts, the reviewer reports each one it
-cannot find in the image as a defect line under `EXPECT_LINE_ID` that quotes it, and
-`reviewed_misses` reads those lines back. Only a render no readable review answered for keeps
-`UNVERIFIED_LINE`.
+`expected_misses` reads matplotlib's drawn texts only. Every other library is checked by the
+reviewer instead (`RuntimeAdapter.review_texts`, owner decision of 2026-10-29): its probe reads
+the texts from the library's own roles or chart model, which a plugin's own drawing is not in,
+so it grounds the reply but never claims a requested text missing. The review request carries
+the requested texts, the reviewer reports each one it cannot find in the image as a defect line
+under `EXPECT_LINE_ID` that quotes it, and `reviewed_misses` reads those lines back. Only a
+render no readable review answered for keeps `UNVERIFIED_LINE`.
 """
 
 import re
@@ -48,6 +54,7 @@
 """One drawn text as the check reads it (`harness.LABEL_CHARS`): longer than any requested text."""
 GROUP_LIMITS: dict[ShownGroup, int] = {
     "title": 3,
+    "subtitle": 2,
     "x axis label": 2,
     "y axis label": 2,
     "z axis label": 2,
@@ -60,13 +67,13 @@
     "other texts": 6,
 }
 """Distinct texts `PlotShows.texts` keeps per role: enough to quote a title or read a tick format, and
-at most 62 texts of `MAX_SHOWN_TEXT_CHARS` (about 5 KB) next to the result's other fields."""
+at most 64 texts of `MAX_SHOWN_TEXT_CHARS` (about 5 KB) next to the result's other fields."""
 UNVERIFIED_LINE = (
-    "SC-04 ({theme}): the requested texts could not be checked, since the renderer reported no drawn texts "
-    "→ compare them with the image. Likely cause: a renderer without the probe's text report."
+    "SC-04 ({theme}): the requested texts could not be checked against the render "
+    "→ compare them with the image. Likely cause: no drawn texts to check them against and no readable review."
 )
-"""The residual line of a render with requested texts but no `labels` in its probe, unless a readable review
-answered for them (`reviewed_misses`): never a silent pass."""
+"""The residual line of a render with requested texts that neither matplotlib's drawn texts nor a readable review
+answered for (`expected_misses`, `reviewed_misses`): never a silent pass."""
 EXPECT_LINE_ID = {"text": "SC-04", "format": "SC-02"}
 """The rubric id a missing requested text (title, legend, label) or number format is reported under."""
 _QUOTED = re.compile(r'["“”„«»]([^"“”„«»\n]+)["“”„«»]')
@@ -118,6 +125,7 @@
 KIND_GROUPS: dict[str, ShownGroup] = {
     "title": "title",
     "suptitle": "title",
+    "subtitle": "subtitle",
     "x axis label": "x axis label",
     "y axis label": "y axis label",
     "z axis label": "z axis label",
@@ -127,7 +135,9 @@
     "z tick label": "z tick labels",
     "annotation": "annotations",
 }
-"""The `PlotShows.texts` role of each probe `kind` (`harness._text_kinds` and `drawn_labels`'s tick labels)."""
+"""The `PlotShows.texts` role of each probe `kind` (`harness._text_kinds` and `drawn_labels`'s tick labels); `subtitle`
+is the text probes' (`harness.text_labels`, `render.mjs`, `text_probe.R`, `text_probe.jl`), a role of its own so a
+subtitle is never trimmed with the other texts."""
 TICK_GROUPS: tuple[ShownGroup, ...] = ("y tick labels", "x tick labels", "z tick labels")
 
 
@@ -194,6 +204,44 @@ def drawn_texts(probe: Any) -> DrawnTexts | None:
     return DrawnTexts(labels, complete)
 
 
+PROBE_ERRORS = frozenset(
+    {
+        "Error",
+        "TimeoutError",
+        "TypeError",
+        "ValueError",
+        "RangeError",
+        "ReferenceError",
+        "ParseError",
+        "KeyError",
+        "IndexError",
+        "AttributeError",
+        "RecursionError",
+        "MemoryError",
+        "OSError",
+        "RuntimeError",
+        "UnicodeDecodeError",
+        "JSONDecodeError",
+        "MethodError",
+        "FieldError",
+        "BoundsError",
+        "UndefVarError",
+        "ArgumentError",
+    }
+)
+"""The exception classes a failed probe names (`{"error": <class>}`: matplotlib's probe, a text probe, Julia's types
+among them; R's probe names `Error`) that the attribution log keeps by name; any other value reads `other`, since the
+code under test can write the probe."""
+
+
+def probe_error(probe: Any) -> str | None:
+    """The class a probe names for a report that failed, for the `pipeline_render` line: None when it names none."""
+    error = probe.get("error") if isinstance(probe, dict) else None
+    if error is None:
+        return None
+    return error if isinstance(error, str) and error in PROBE_ERRORS else "other"
+
+
 def plot_texts(drawn: DrawnTexts | None) -> dict[ShownGroup, list[str]] | None:
     """`PlotShows.texts`: the drawn texts within `GROUP_LIMITS` and `MAX_SHOWN_TEXT_CHARS`, for the root's reply."""
     if drawn is None: