@@ -1505,12 +1505,34 @@ npm run deploy
     not one on these. Do not re-open it on the size table alone: that table now
     favours dcb, and it is the wrong axis.
 
-    One cost the byte comparison hides: `node:zlib`'s brotli has NO dictionary
-    parameter (MODE, QUALITY, LGWIN, LGBLOCK, DISABLE_LITERAL_CONTEXT_MODELING,
-    SIZE_HINT, LARGE_WINDOW, NPOSTFIX, NDIRECT is the whole list). Those dcb
-    figures came from shelling out to the `brotli` CLI with `-D`. Switching would
-    reintroduce a CLI dependency in the build path, which is precisely what moving
-    the deltas into the build deleted.
+    One cost the byte comparison used to hide is RETIRED, and it is worth reading
+    as a correction rather than a fact. This note said `node:zlib`'s brotli has NO
+    dictionary parameter, listed the nine params as the whole list, and concluded
+    that switching would reintroduce a `brotli` CLI dependency in the build path,
+    which is precisely what moving the deltas into the build deleted. Node 26 takes
+    a `dictionary` option (nodejs/node#61763, merged 2026-02-13), and
+    `.node-version` is 26, so CI has it too. Measured 2026-08-10 on node v26.7.0,
+    q11 over a 21KB target: 70 bytes with no dictionary against 18 with one. The
+    dcb figures above still came from the CLI with `-D`, because that was the only
+    door on the day they were taken.
+
+    The CLI-dependency argument is therefore gone while the CONCLUSION is
+    untouched, since dcz won on DECODE and neither engine's encoder was ever the
+    question. Re-open it only with a fresh decode measurement, never on the size
+    table (see above) and no longer on this paragraph.
+
+    **Probe an engine's dictionary support, never infer it.** Three runtimes
+    disagree on this one option. Node 26 honours it. workerd accepts it for zstd
+    and silently ignores it, measured 2026-08-05, which is why `/terminal` ships no
+    delta calculator and says so on the page. Bun does the same thing through
+    1.3.14, its current stable, with the fix landing 2026-07-18 in oven-sh/bun#34427
+    and verified here on `1.4.0-canary.1`. **The failure is silent in every case,
+    because a frame compressed WITHOUT the dictionary still decodes fine WITH it**,
+    so nothing throws and the only signal is a byte count that never shrank. The
+    control is four lines: compress one target with no dictionary, the right one,
+    and a wrong one. An engine that honours the option prints a smaller number for
+    the right dictionary alone, and an engine that ignores it prints the same
+    number three times.
 
     **zstd above level 19 is dead weight here.** Levels 19, 20, 21, 22, 22+long-
     distance-matching and 22+btultra2 produce BYTE-IDENTICAL output on all 9 shell
@@ -1527,8 +1549,9 @@ npm run deploy
 
     Deltas are BUILD OUTPUT, generated by build.mjs with `node:zlib`'s zstd. An
     earlier version of this note said dictionary compression was "unreachable from
-    Node" and shipped a workstation script with committed artifacts; that limit is
-    BROTLI's, and generalizing it to zstd was wrong. `zstdCompressSync` takes a
+    Node" and shipped a workstation script with committed artifacts; that limit was
+    BROTLI's, generalizing it to zstd was wrong, and node 26 has since retired the
+    brotli half of it too (see above). `zstdCompressSync` takes a
     `dictionary` option, it beats shelling out (116 bytes where the zstd CLI gave
     120), and the foreign `zstd -d -D` CLI decodes Node's output byte-exact,
     skippable prefix included — the interop check that matters, since the real
 
@@ -1595,8 +1595,10 @@ for (const [file, srcPath, marker] of SHELLS) {
   // luna.css change: 116 bytes against 7,615.
   //
   // This used to be a workstation script with committed artifacts, on the belief that
-  // dictionary compression was unreachable from Node. That was true of BROTLI and I wrongly
-  // generalized it: node:zlib's zstd DOES take a `dictionary` option. It is also better
+  // dictionary compression was unreachable from Node. That was true of BROTLI AT THE TIME and I
+  // wrongly generalized it: node:zlib's zstd DOES take a `dictionary` option. (node 26 has since
+  // given brotli one too, nodejs/node#61763, which retires the limit rather than the choice —
+  // dcz won on decode speed, not on either encoder's availability. See gotcha 14.) It is also better
   // than shelling out — 116 bytes where the zstd CLI produced 120 — and portable, verified
   // by having the foreign `zstd -d -D` CLI decode Node's bytes byte-exact, skippable prefix
   // and all. That interop check is the one that matters, because the real decoder is a
 
@@ -379,24 +379,36 @@ <h2>Deltas, and why they ship as dcz</h2>
     </p>
     <p>
       RFC 9842 defines two encodings for this, and Cloudflare passes both through identically, so
-      the choice is free and purely about engineering. Brotli is nominally smaller. zstd decodes
-      about twice as fast:
+      the choice is free and purely about engineering. Brotli compresses smaller. zstd decodes
+      much faster:
     </p>
     <table class="cmp">
       <thead><tr><th>measure</th><th>dcb (brotli)</th><th>dcz (zstd)</th></tr></thead>
       <tbody>
-        <tr><td>the real luna.css delta</td><td>79 bytes</td><td>80 bytes</td></tr>
-        <tr><td>source-level deltas, larger</td><td class="y">5-8% smaller</td><td>n/a</td></tr>
-        <tr><td>decode, bare 46KB asset</td><td>0.094 ms</td><td class="y">0.046 ms</td></tr>
-        <tr><td>throughput</td><td>471 MB/s</td><td class="y">954 MB/s</td></tr>
+        <tr><td>all 12 shipping deltas, total</td><td class="y">3,344 bytes</td><td>3,589 bytes</td></tr>
+        <tr><td>widest single gap (nav.js)</td><td class="y">685 bytes</td><td>729 bytes</td></tr>
+        <tr><td>decode, 47,615 B rebuilt</td><td>0.1368 ms</td><td class="y">0.0165 ms</td></tr>
+        <tr><td>throughput</td><td>332 MiB/s</td><td class="y">2,752 MiB/s</td></tr>
       </tbody>
     </table>
     <p>
-      I argued for brotli on bytes and got talked out of it, correctly. Bytes are a proxy;
-      latency is the thing you actually want. And the byte case collapses on inspection: on the
-      pair this site really ships, brotli wins by <em>one byte</em>. The 5-8% edge only appears
-      on much larger deltas than a shell asset produces. Meanwhile the decode numbers came off a
-      fast laptop, so the gap widens on the phones that need it most.
+      I argued for brotli on bytes and got talked out of it, correctly, and then the numbers I
+      argued from turned out to be wrong in <em>both</em> directions. Re-measured against real
+      dictionaries rather than stand-ins, brotli's size edge is bigger than the one byte I had
+      (245 bytes across the set, 6.8%, winning 11 of 12), and zstd's decode edge is far bigger
+      than the 2x I had (8.3x). The call held anyway, which is the interesting part.
+    </p>
+    <p>
+      It held for a structural reason: <strong>decode scales with the reconstruction, not with
+      the delta.</strong> A 685-byte dcb delta still rebuilds 47,615 bytes, so making the delta
+      smaller never makes the decode gap smaller. Break-even at 9 Mbps is about 135 bytes saved
+      per asset and dcb saves 44, so dcz wins by roughly 3x. On a slow phone the margin widens,
+      because decode scales with the CPU while 44 bytes stays 44 bytes.
+    </p>
+    <p>
+      Worth saying plainly: on the old numbers this was a coin flip, and on these it is not. The
+      size table now favours brotli by more than it ever did, and that table is the wrong axis to
+      decide on.
     </p>
     <p>
       dcz's framing is the tidier of the two as a bonus. The dictionary hash rides in a Zstandard
@@ -422,7 +434,7 @@ <h2>Deltas, and why they ship as dcz</h2>
     <h2>One more over-generalization</h2>
     <p>
       Dictionary compression is unreachable from Node's zlib, so the deltas started life as a
-      workstation script with committed artifacts. That sentence is true of brotli. I applied it
+      workstation script with committed artifacts. That sentence was true of brotli. I applied it
       to zstd without checking, and <code>zstdCompressSync</code> takes a <code>dictionary</code>
       option. It also beats shelling out to the CLI: 116 bytes against 120.
     </p>
@@ -433,6 +445,15 @@ <h2>One more over-generalization</h2>
       interop: the foreign <code>zstd -d -D</code> CLI decodes Node's output byte-exact, prefix
       included, because the real decoder is a browser and not Node.
     </p>
+    <p>
+      <strong>Update, August 2026:</strong> the brotli half expired too. Node 26 added a
+      <code>dictionary</code> option to brotli
+      (<a href="https://github.com/nodejs/node/pull/61763" rel="external">nodejs/node#61763</a>),
+      so on v26.7.0 a quality-11 compress of a 21 KB target goes from 70 bytes with no dictionary
+      to 18 with one. The limit I over-generalized was accurate about brotli the day I wrote it
+      and is now accurate about neither encoder. So a runtime limit has a shelf life, and the
+      version you measured on belongs in the note next to the measurement.
+    </p>
 
     <h2>What shipped</h2>
     <table class="cmp">
@@ -511,17 +532,17 @@ <h2>The part that transfers</h2>
       "q": "The deltas ship as dcz even though brotli measured smaller. Why?",
       "options": [
         {
-          "t": "zstd decodes about 2x faster, and on the pair this site actually ships brotli's advantage was one byte.",
+          "t": "Decode scales with the reconstruction rather than the delta, so brotli's smaller delta never shrinks zstd's 8.3x decode lead.",
           "ok": true,
-          "why": "Bytes are a proxy for latency. The 5-8% brotli edge only shows up on deltas far larger than a shell asset produces, while the decode gap grows on slow CPUs."
+          "why": "Re-measured against real dictionaries, dcb is 245 bytes (6.8%) ahead across all 12 pairs, and dcz still wins: rebuilding 47,615 bytes costs 0.0165 ms against 0.1368. Break-even at 9 Mbps needs ~135 bytes saved per asset and dcb saves 44."
         },
         {
           "t": "Cloudflare only passes dcz through; dcb is not supported at the edge.",
           "why": "Both are passed through identically, on all plans. That is exactly why the choice was free to make on engineering grounds."
         },
         {
           "t": "Because brotli cannot use a custom dictionary, so dcb was never actually available.",
-          "why": "Brotli does custom dictionaries fine via its CLI. What it lacks is a dictionary option in Node's zlib, which is a tooling gap rather than a format limit."
+          "why": "Brotli does custom dictionaries fine. What it lacked was a dictionary option in Node's zlib, which was a tooling gap rather than a format limit, and node 26 closed even that."
         }
       ]
     },
 
@@ -5,7 +5,8 @@
 //
 // This is all that remains of the old gen-shell-deltas.mjs. The deltas themselves moved
 // into build.mjs once it turned out node:zlib's zstd takes a `dictionary` option (the
-// "unreachable from Node" limit was BROTLI's, and I had over-generalized it), so they are
+// "unreachable from Node" limit was BROTLI's at the time, I had over-generalized it, and
+// node 26 has since given brotli one as well), so they are
 // build output now: no zstd CLI, no committed artifacts, nothing to forget.
 //
 // Rolling stays a deliberate human step for one reason: it WRITES INTO THE SOURCE TREE, and