docs: make the plan agree with the measurement it already records

Review found CHAT-TYPOGRAPHY.md contradicting itself in three places. One cause:
D1 was corrected after measuring, and the correction was never propagated.

  intro            "~130-character lines"  → ~154, the figure actually measured
  T1 exit          "~72 characters"        → ~116, the 680px cap D1 settled on
  risks            "no test guards this"   → closed by the guards in this PR

The 130 and the 72 were both written BEFORE the measurement that produced 680px;
the risk bullet was true when written and stopped being true in the same commit.

Rather than proofreading it again, the number is now pinned to the code: a test
reads `--prose-max` out of the stylesheet and asserts the doc names that exact
cap, and that no stray reference to the disproved 72-character target survives.
Prose drifts from code silently — this is the only claim in the doc that a test
can hold, so it holds it. Caught both directions: change the CSS without the doc
(15/18), the doc without the CSS (15/18), reintroduce the old figure (15/18).

The fourth comment (detachedHtml has no CSP) was written against this branch
before it was rebased, when it still carried #70's commits. extension.js is no
longer in this PR's diff, and the CSP + nonce landed in #70 as dc91502.

33 suites green.

Author: ndemianc

docs/CHAT-TYPOGRAPHY.md

@@ -10,7 +10,7 @@ activity. This doc takes the *visual layer*, which that one deliberately left al
 
 This became urgent the moment the chat could open as an editor tab (#70). In a 380px sidebar the
 line length is bounded by the container, so nothing looks badly wrong. At 900px it is unbounded,
-and the same CSS produces ~130-character lines — roughly twice the readable measure.
+and the same CSS produced **~154-character lines** — measured, at a 900px viewport.
 
 ---
 
@@ -130,8 +130,9 @@ Each ships independently and is visible on its own.
 
 **T1 — measure + rhythm** *(S)*. D1 and D3. The largest perceptual change for the least code, and the
 one that fixes the editor tab. Ships: a wrapper max-width, `em`-based prose spacing, wider `#log`
-padding at editor width. **Exit:** a long answer in the editor tab holds ~72 characters per line, and
-the sidebar renders byte-identically to today.
+padding at editor width. **Exit:** a long answer in the editor tab holds **~116 characters** per line
+(the 680px cap of D1, down from ~154), and the sidebar renders byte-identically to today — verified by
+comparing computed styles against `develop` at 520px, not by eye.
 
 **T2 — the reading type scale** *(S)*. D2 and D4. Ships: the prose size/leading custom properties and
 the widened heading scale. **Exit:** h1/h2/h3 are distinguishable at a glance in a screenshot with no
@@ -157,9 +158,11 @@ by itself is the cheapest way to find out before spending effort on T2–T4.
 - **Theme variance.** Inline-code colour already comes from the theme rather than from us (§1), so
   any judgement about "busy" colour must be checked across the light, dark and high-contrast themes
   the `webviewCss` suite already reasons about — not just the default.
-- **No test currently guards the measure.** `webviewCss.test.js` pins hidden-attribute defeats and the
-  session-card overflow; it should gain a guard that the prose column is bounded, or T1 will regress
-  silently the first time someone refactors the log container.
+- **~~No test currently guards the measure.~~ Closed by T1.** `webviewCss.test.js` now asserts the
+  column is bounded, that the cap covers every child of the log rather than just `.msg`, that it stays
+  an absolute length (not `ch`), and that the rhythm change stays width-gated. The risk was real: this
+  regression is invisible in a sidebar, so whoever refactors the log container would not see it break —
+  a user with the chat in an editor tab would.
 - **Screenshots are not measurements.** Everything here is derived from our own CSS plus a
   side-by-side comparison. The specific numbers (72ch, 14px, 1.65) are considered starting points to
   be tuned against the real thing at real widths, not values copied from the reference — we cannot

extensions/levelcode-ai/test/webviewCss.test.js

@@ -268,6 +268,22 @@ test('TRANSCRIPT: the prose column is bounded, and every child shares the one me
 		'the measure should be an absolute length, not `ch` — see CHAT-TYPOGRAPHY.md D1');
 });
 
+test('TRANSCRIPT: the design doc quotes the SAME measure the stylesheet ships', () => {
+	// Review found the doc contradicting itself in three places: the cap was corrected in D1 after
+	// measuring, and the intro, the exit criterion and the risks list kept the pre-measurement numbers.
+	// Prose drifts from code silently, so the one number that matters is pinned to the code instead of
+	// to a proofread.
+	const doc = fs.readFileSync(path.join(__dirname, '..', '..', '..', 'docs', 'CHAT-TYPOGRAPHY.md'), 'utf8');
+	const shipped = /--prose-max:\s*(\d+)px/.exec(css);
+	assert.ok(shipped, 'the stylesheet no longer declares --prose-max');
+
+	assert.ok(doc.includes('**Target ' + shipped[1] + 'px**'),
+		'CHAT-TYPOGRAPHY.md D1 does not name the ' + shipped[1] + 'px cap the stylesheet actually ships');
+	// And no stray reference to the estimate that measuring disproved.
+	assert.ok(!/~72 characters per line/.test(doc),
+		'the doc still quotes the pre-measurement 72-character target somewhere');
+});
+
 test('TRANSCRIPT: the looser rhythm is gated to reading width, so the sidebar is untouched', () => {
 	// T1's exit criterion is that a narrow panel renders exactly as before — a user who upgrades and
 	// never opens the editor tab should see nothing move. Verified against develop's computed styles