diff --git a/architecture.md b/architecture.md
index 41a715b92..221423a00 100644
--- a/architecture.md
+++ b/architecture.md
@@ -7085,6 +7085,19 @@ Batch-local ids keep that set as small as one batch. This bounds the request,
 not the document: nothing is truncated to make an engine accept it, and the
 answers come back one per candidate in the order they were asked for.
 
+**An engine bounds the text one render measures as well as its elements.** Every
+space-separated part of every text in a render is kept in a cache the engine
+refuses from, and that cache holds the same number of them however few elements
+carry them — so a batch well inside the element budget still overruns it when
+its rows are prose, and one line of a retained reading can hold more parts than
+any render measures whatever a batch size is. A batch is therefore bounded by
+both, and a line is measured as far as it could still fit its region rather than
+to its end: every part after the first is preceded by a space, so a prefix of
+more parts than the region has columns is already wider than the region, and
+text does not narrow when more of it is appended. The cut that prefix bounds is
+the reader's own line, measured as the row it became — a bounded question, not a
+shortened document.
+
 An engine that is asked for too much at once does not say so usefully. It
 reports unbalanced structure for a list whose opens and closes are exactly
 paired, which is not a capacity refusal — so it is not reclassified as one
diff --git a/packages/cli/src/repl/fitting.ts b/packages/cli/src/repl/fitting.ts
index 6e3902237..4c5e0c8fb 100644
--- a/packages/cli/src/repl/fitting.ts
+++ b/packages/cli/src/repl/fitting.ts
@@ -306,6 +306,61 @@ const SEPARATOR = "  ";
  */
 const WIDTHS_PER_RENDER = 1024;
 
+/**
+ * How many space-separated parts one measuring render may carry.
+ *
+ * The element budget is not the engine's only one. Every part of every text a
+ * render measures is kept in a cache the engine refuses from, and that cache
+ * holds the same number of them however few elements carry them: measured on a
+ * freshly built engine, one text of 16,384 parts and 1,366 texts of twelve each
+ * refuse, while 16,383 parts go through either way. A tab and a newline are not
+ * separators there, a run of spaces produces an empty part that costs what a
+ * word costs, and two texts spelled the same share one entry.
+ *
+ * This is the budget whose refusal the renderer's recovery cannot absorb: a
+ * replacement instance is given the same cache, so a render that overran it
+ * overruns the replacement too and the refusal reaches the screen as the error
+ * that ends it. Half the engine's own figure, which leaves the parts an
+ * instance is still holding from the render before this one somewhere to sit.
+ */
+const WORDS_PER_RENDER = 8192;
+
+/** What one text costs that budget, counted the way the engine counts it. */
+function wordsIn(text: string): number {
+  let parts = 1;
+  for (let at = 0; at < text.length; at += 1) {
+    if (text[at] === " ") {
+      parts += 1;
+    }
+  }
+  return parts;
+}
+
+/**
+ * How much of one line could still fit a region, as a length in code units.
+ *
+ * A single text is bounded by no batch size, and one line of somebody's output
+ * can hold more parts than a render measures — a journal of this product holds
+ * lines of 19,999 — so a line is measured as far as it could still fit and
+ * never further. That prefix settles the question it is asked: every part after
+ * the first is preceded by a space, so `room + 2` of them are already wider
+ * than `room` columns, and text does not narrow when more of it is appended.
+ * The prefix ends at a space, which is a grapheme boundary of its own.
+ */
+function windowOf(text: string, room: number): number {
+  const parts = Math.min(room + 2, WORDS_PER_RENDER);
+  let seen = 0;
+  for (let at = 0; at < text.length; at += 1) {
+    if (text[at] === " ") {
+      seen += 1;
+      if (seen === parts) {
+        return at;
+      }
+    }
+  }
+  return text.length;
+}
+
 /**
  * Ask the engine how wide each of these texts is.
  *
@@ -324,12 +379,26 @@ function* widths(
     return Ok([]);
   }
   const answers: number[] = [];
-  for (let from = 0; from < texts.length; from += WIDTHS_PER_RENDER) {
-    const batch = yield* batchWidths(renderer, size, texts.slice(from, from + WIDTHS_PER_RENDER));
+  let from = 0;
+  while (from < texts.length) {
+    let to = from;
+    let words = 0;
+    while (to < texts.length && to - from < WIDTHS_PER_RENDER) {
+      const cost = wordsIn(texts[to]);
+      // A text that outgrows the budget on its own still goes, alone, so the
+      // engine's refusal names that text rather than this loop never advancing.
+      if (to > from && words + cost > WORDS_PER_RENDER) {
+        break;
+      }
+      words += cost;
+      to += 1;
+    }
+    const batch = yield* batchWidths(renderer, size, texts.slice(from, to));
     if (!batch.ok) {
       return batch;
     }
     answers.push(...batch.value);
+    from = to;
   }
   return Ok(answers);
 }
@@ -679,11 +748,15 @@ function* fitLine(
   if (line.text.length === 0) {
     return Ok([piece(0, 0, false, 0)]);
   }
-  const whole = yield* widths(renderer, size, [line.text]);
-  if (!whole.ok) {
-    return whole;
+  // Asked of as much of the line as the region could still hold. One row is a
+  // line the engine measured as fitting, and a line held back from the question
+  // is longer than the answer it would have given, so it is cut.
+  const reach = windowOf(line.text, room);
+  const reaching = yield* widths(renderer, size, [line.text.slice(0, reach)]);
+  if (!reaching.ok) {
+    return reaching;
   }
-  if (whole.value[0] <= room) {
+  if (reach === line.text.length && reaching.value[0] <= room) {
     return Ok([piece(0, line.text.length, false, 0)]);
   }
   // Where this line begins, so its continuations begin there too. Bounded well
@@ -728,7 +801,10 @@ function* cutAt(
   rest: string,
   room: number,
 ): Operation<Result<number>> {
-  const stops = graphemeStops(rest);
+  // Over the boundaries of the prefix that could still fit, which is what keeps
+  // every candidate inside one render's budget and keeps this from segmenting
+  // the whole remainder once per row of a very long line.
+  const stops = graphemeStops(rest.slice(0, windowOf(rest, room)));
   // Narrow by halves over grapheme boundaries, asking the engine for one batch
   // of candidates per round rather than one render per candidate.
   let low = 0;
diff --git a/packages/cli/tests/repl-source-reading.test.ts b/packages/cli/tests/repl-source-reading.test.ts
index 4b845611a..a8045ab61 100644
--- a/packages/cli/tests/repl-source-reading.test.ts
+++ b/packages/cli/tests/repl-source-reading.test.ts
@@ -1176,3 +1176,90 @@ describe("S6 a reading longer than one measuring render is still measured whole"
       }));
   }
 });
+
+/**
+ * S7 a reading with more words than one measuring render.
+ *
+ * The engine's element budget is not its only one. It keeps every
+ * space-separated part of every text a render measures in a cache it refuses
+ * from, and that cache holds 16,384 of them however few elements carry them —
+ * measured on a freshly built engine, 1,366 texts of twelve parts and one text
+ * of 16,384 both refuse. A refusal a bigger engine cannot fix is the one the
+ * renderer's recovery cannot absorb, so it reached the screen as the error that
+ * ends it.
+ *
+ * Both arms are the shape the reported crash had. An entry whose recorded
+ * output holds one line of 20,000 parts is the real one — a journal of this
+ * product holds lines of 19,999 — and it refused inside the first measurement
+ * of that single line, which no batch size bounds. A reading of 1,400 ordinary
+ * lines is the other half: each line measures on its own, and the pass that
+ * measures every finished row at once asked for more parts than the cache holds
+ * even though it asked for fewer elements than the element budget.
+ */
+/** One line of `count` space-separated parts, which is what the engine counts. */
+function parts(count: number): string {
+  return Array.from({ length: count }, (_, at) => `${at % 10}`).join(" ");
+}
+
+describe("S7 a reading with more words than one measuring render is still measured whole", () => {
+  /** One recorded output line, as the reading an entry is shown as. */
+  function outputLinesOf(output: string): readonly ReplReadingLine[] {
+    return entryReading({
+      entry: entryOf('<Json value={{ a: 1 }} as="a" />', {
+        terminal: settled("ok", output),
+        settled: true,
+      }),
+      lifecycle: NO_LIFECYCLE,
+      live: "",
+      inspected: false,
+    }).output.lines;
+  }
+
+  it("recovers one line of 20,000 parts, exactly and in order", () =>
+    withRenderer(function* (renderer) {
+      const line = parts(20000);
+      const lines = outputLinesOf(line);
+      const prepared = yield* prepareReading(renderer, WIDE, lines, 100);
+      if (!prepared.ok) {
+        throw prepared.error;
+      }
+
+      // One line of somebody's output, whole: it was wrapped into many rows and
+      // nothing was shortened to make the engine accept it.
+      const whole = recovered(prepared.value);
+      expect(whole.size).toBe(lines.length);
+      for (const one of lines) {
+        expect([one.key, whole.get(one.key)]).toEqual([one.key, one.text]);
+      }
+      const rows = prepared.value.rows.filter((one) => one.line === lines[lines.length - 1].key);
+      expect(rows.length).toBeGreaterThan(1);
+
+      // And every row still fits the region it was cut for, which is the claim a
+      // width answered for a bounded prefix could break.
+      const { reservation } = prepared.value;
+      for (const row of prepared.value.rows) {
+        expect(row.width).toBeLessThanOrEqual(reservation.content - row.indent);
+        expect(runText(row.runs)).toBe(row.text);
+      }
+    }));
+
+  it("recovers 1,400 ordinary lines whose parts outgrow one render", () =>
+    withRenderer(function* (renderer) {
+      // Distinct lines, because the engine caches one entry per exact text: a
+      // thousand copies of one line cost it the parts of a single line and
+      // would prove nothing about what a batch of them asks for.
+      const lines = outputLinesOf(
+        Array.from({ length: 1400 }, (_, at) => `line ${at} ${parts(20)}`).join("\n"),
+      );
+      const prepared = yield* prepareReading(renderer, WIDE, lines, 100);
+      if (!prepared.ok) {
+        throw prepared.error;
+      }
+
+      const whole = recovered(prepared.value);
+      expect(whole.size).toBe(lines.length);
+      for (const one of lines) {
+        expect([one.key, whole.get(one.key)]).toEqual([one.key, one.text]);
+      }
+    }));
+});