chore(registry): record measured CPU throughput for all seven models

Every model said expected_time_per_1k_texts_cpu: benchmark_required,
not just the four PsyEmbedding ones. Measured with
scripts/bench_models.py --threads 2 --quick.

Throughput varies about 50x between a 15-word social post and a
250-word document, so the field is three numbers rather than one, and
the comment names the host and date: a shared cloud vCPU runs roughly
2-4x slower than the machine these came from, so they need a re-run on
the Space before they are shown to users.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

packages/model_registry/models.yaml

@@ -20,7 +20,14 @@ models:
       pooling: sentence_transformers_default   # descriptive only - never reimplement
       normalize_embeddings: true
     operational_config:
-      expected_time_per_1k_texts_cpu: benchmark_required
+      # Measured 2026-08-26: scripts/bench_models.py --threads 2 --quick on an
+      # Apple M4 Pro. Throughput is dominated by per-core CPU speed, so a shared
+      # cloud vCPU runs roughly 2-4x slower; re-run on the host before quoting
+      # these to users. Seconds per 1,000 texts, by text length.
+      expected_time_per_1k_texts_cpu:
+        short_15w_s: 0.8
+        medium_60w_s: 1.0
+        long_250w_s: 2.8
       worker_ram_min_gb: 2
       lazy_load: false
     warnings: []
@@ -43,7 +50,14 @@ models:
       pooling: average_pool                    # descriptive only - never reimplement
       normalize_embeddings: true
     operational_config:
-      expected_time_per_1k_texts_cpu: benchmark_required
+      # Measured 2026-08-26: scripts/bench_models.py --threads 2 --quick on an
+      # Apple M4 Pro. Throughput is dominated by per-core CPU speed, so a shared
+      # cloud vCPU runs roughly 2-4x slower; re-run on the host before quoting
+      # these to users. Seconds per 1,000 texts, by text length.
+      expected_time_per_1k_texts_cpu:
+        short_15w_s: 2.5
+        medium_60w_s: 9.0
+        long_250w_s: 40.5
       worker_ram_min_gb: 4
       worker_ram_recommended_gb: 8
       lazy_load: true
@@ -70,7 +84,14 @@ models:
       pooling: average_pool                    # descriptive only - never reimplement
       normalize_embeddings: true
     operational_config:
-      expected_time_per_1k_texts_cpu: benchmark_required
+      # Measured 2026-08-26: scripts/bench_models.py --threads 2 --quick on an
+      # Apple M4 Pro. Throughput is dominated by per-core CPU speed, so a shared
+      # cloud vCPU runs roughly 2-4x slower; re-run on the host before quoting
+      # these to users. Seconds per 1,000 texts, by text length.
+      expected_time_per_1k_texts_cpu:
+        short_15w_s: 1.3
+        medium_60w_s: 3.3
+        long_250w_s: 14.7
       worker_ram_min_gb: 4
       lazy_load: true
       max_resident_models: 1
@@ -99,7 +120,14 @@ models:
       pooling: mean                            # per model card
       normalize_embeddings: true
     operational_config:
-      expected_time_per_1k_texts_cpu: benchmark_required
+      # Measured 2026-08-26: scripts/bench_models.py --threads 2 --quick on an
+      # Apple M4 Pro. Throughput is dominated by per-core CPU speed, so a shared
+      # cloud vCPU runs roughly 2-4x slower; re-run on the host before quoting
+      # these to users. Seconds per 1,000 texts, by text length.
+      expected_time_per_1k_texts_cpu:
+        short_15w_s: 2.5
+        medium_60w_s: 8.9
+        long_250w_s: 40.6
       worker_ram_min_gb: 4
       lazy_load: true
       max_resident_models: 1
@@ -124,7 +152,14 @@ models:
       pooling: mean                            # per model card
       normalize_embeddings: true
     operational_config:
-      expected_time_per_1k_texts_cpu: benchmark_required
+      # Measured 2026-08-26: scripts/bench_models.py --threads 2 --quick on an
+      # Apple M4 Pro. Throughput is dominated by per-core CPU speed, so a shared
+      # cloud vCPU runs roughly 2-4x slower; re-run on the host before quoting
+      # these to users. Seconds per 1,000 texts, by text length.
+      expected_time_per_1k_texts_cpu:
+        short_15w_s: 2.7
+        medium_60w_s: 9.1
+        long_250w_s: 41.5
       worker_ram_min_gb: 4
       lazy_load: true
       max_resident_models: 1
@@ -149,7 +184,14 @@ models:
       pooling: mean                            # per model card
       normalize_embeddings: true
     operational_config:
-      expected_time_per_1k_texts_cpu: benchmark_required
+      # Measured 2026-08-26: scripts/bench_models.py --threads 2 --quick on an
+      # Apple M4 Pro. Throughput is dominated by per-core CPU speed, so a shared
+      # cloud vCPU runs roughly 2-4x slower; re-run on the host before quoting
+      # these to users. Seconds per 1,000 texts, by text length.
+      expected_time_per_1k_texts_cpu:
+        short_15w_s: 2.5
+        medium_60w_s: 9.0
+        long_250w_s: 42.0
       worker_ram_min_gb: 4
       lazy_load: true
       max_resident_models: 1
@@ -177,7 +219,14 @@ models:
       pooling: mean                            # per model card
       normalize_embeddings: true
     operational_config:
-      expected_time_per_1k_texts_cpu: benchmark_required
+      # Measured 2026-08-26: scripts/bench_models.py --threads 2 --quick on an
+      # Apple M4 Pro. Throughput is dominated by per-core CPU speed, so a shared
+      # cloud vCPU runs roughly 2-4x slower; re-run on the host before quoting
+      # these to users. Seconds per 1,000 texts, by text length.
+      expected_time_per_1k_texts_cpu:
+        short_15w_s: 2.5
+        medium_60w_s: 8.9
+        long_250w_s: 40.7
       worker_ram_min_gb: 4
       lazy_load: true
       max_resident_models: 1