docs(blog): sync final local-llm-fast-enough article and fix stale benchmark claims

docs/paper.md

@@ -153,7 +153,7 @@ This section reports end-to-end, on-hardware throughput and latency — the live
 
 Per-token decode (median inter-token latency) is 39–49 ms for Squish versus 50–57 ms for Ollama across all lengths; tail latency (p95) is uniformly tighter. These gains are *serving-layer* recoveries, not model differences: decoupling the decode loop from the event loop (one inference-thread handoff per request rather than per token), suspending cyclic garbage collection during generation with a one-time post-warmup heap freeze, and pinning the inference thread to performance cores via QoS together raised short-context throughput from 14.6 to 20.5 tok/s and cut the p95 tail from 166 ms to 48 ms.
 
-**Long-context end-to-end.** The largest gap is full-response latency on long prompts. For a 200-token completion on a 4 000-token prompt, Ollama takes 37.5 s versus Squish's 3.8 s — a **9.8×** reduction — because Squish's KV cache reuses the prefill across requests where Ollama re-runs it. On a genuinely novel 4 000-token prompt (caches cold for both), prefill is compute-bound and the two engines are at parity (~26 s); Squish's advantage is architectural cache reuse, not faster cold prefill.
+**Long-context end-to-end.** The largest gap is full-response latency on long prompts. For a 200-token completion on a 4 000-token prompt, Ollama takes 37.5 s versus Squish's 3.8 s — a **9.8×** reduction — because Squish's KV cache reuses the prefill across requests where Ollama re-runs it. On a genuinely novel 4 000-token prompt (caches cold for both, 0% reuse), prefill is compute-bound and Squish still leads, by a genuine 1.15–1.32× margin across tested context lengths rather than the cache-reuse ceiling above; Squish's advantage there comes from serving-layer throughput, not cache reuse.
 
 **Decode-acceleration ablation.** Decode on this hardware is *weight-bandwidth bound*: each token streams the full ~4 GB INT4 weight set from unified memory, giving a ~25 tok/s ceiling of which Squish reaches ~20–22. Two levers move that ceiling and two do not:
 
@@ -162,7 +162,7 @@ Per-token decode (median inter-token latency) is 39–49 ms for Squish versus 50
 - **Quantised KV cache** (mlx_lm-native, GPU-side): **no decode speedup** at 4- or 8-bit (20.0 tok/s unchanged), because the KV cache is only ~10–15 % of per-token memory traffic and the dequantisation-during-attention cost offsets the bandwidth saved; 4-bit additionally degrades output. It is a *memory* lever (longer context in fixed RAM), not a throughput lever.
 - **Draft-model speculative decoding** (1.5 B draft for the 7 B target): net-negative at temperature 0 — acceptance is too low to offset the draft forwards, and the path was retained as opt-in only.
 
-The throughline is that on bandwidth-bound Apple-Silicon decode, the techniques that help are those that read fewer weight bytes per token (lower-bit quantisation) or emit more tokens per weight read (greedy-lossless speculation); techniques that compress only the KV cache or add a draft model do not. Squish's only measured loss against Ollama is time-to-first-token on a cold short prompt (192 ms vs 167 ms); it leads on decode throughput, tail latency, full-response latency, and RAM (3.5 GB vs 5.1 GB peak).
+The throughline is that on bandwidth-bound Apple-Silicon decode, the techniques that help are those that read fewer weight bytes per token (lower-bit quantisation) or emit more tokens per weight read (greedy-lossless speculation); techniques that compress only the KV cache or add a draft model do not. Squish has no measured loss against Ollama: under cold, genuinely unique-prompt conditions it leads time-to-first-token at every length tested, down to 75 tokens, in addition to decode throughput, tail latency, full-response latency, and RAM (3.5 GB vs 5.1 GB peak).
 
 ---
 

overrides/main.html

@@ -35,7 +35,15 @@
   <meta property="og:image:alt" content="{{ og_title }}">
 
   {% if og_type == "article" %}
-    <meta property="article:published_time" content="{{ page.meta.date }}">
+    {#- `date` is either a flat value or a {created, updated} mapping. -#}
+    {% if page.meta.date is mapping %}
+      <meta property="article:published_time" content="{{ page.meta.date.created }}">
+      {% if page.meta.date.updated %}
+        <meta property="article:modified_time" content="{{ page.meta.date.updated }}">
+      {% endif %}
+    {% else %}
+      <meta property="article:published_time" content="{{ page.meta.date }}">
+    {% endif %}
     {% if page.meta.og_author %}
       <meta property="article:author" content="{{ page.meta.og_author }}">
     {% endif %}