fix(providers): register glm-5.2 — the model the published numbers came from

A bench-matrix run naming glm-5.2 failed with 'Unknown provider' and wrote an
error cell: only glm-4.6 and glm-5.1 were in the catalog, so the model behind
this repo's flagship scorecard could not be re-run through its own CLI.

Found while re-running mbpp-plus under the newly-wired plus harness. The
fallback to glm-5.1 CONFOUNDED the result — grading changed and the model
changed at once, so the 99.7% -> 85.4% drop could not be attributed. Registering
5.2 makes the re-run a one-variable measurement against the published number.

Verified served by the same z.ai Anthropic-compat endpoint as 5.1; bare id and
the [1m] long-context tag both answer. Cost row mirrors 5.1 and matches
PRICING, so the catalog-vs-cost.py parity gate stays green (glm-5.2 remains a
declared PRICING_PLACEHOLDER pending Zhipu's public sheet).

Gates: 1784 passed / 25 skipped, ruff clean, mypy 731 files.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

Author: 0bserver07

CHANGELOG.md

@@ -9,6 +9,17 @@ commit receipts.
 
 ### Fixed
 
+- **`glm-5.2` was missing from the provider catalog** — the model this repo
+  publishes its flagship scorecard with. A `bench-matrix --model glm-5.2` run
+  failed with `Unknown provider: 'glm-5.2'` and recorded an **error cell**, so
+  the model behind the published numbers could not be re-run through the CLI at
+  all; only `glm-4.6` and `glm-5.1` were registered. Found while re-running the
+  mbpp-plus column: the fallback to `glm-5.1` silently confounded the
+  comparison, because the grading change and a model change would have moved
+  the number together. Verified served by the same z.ai Anthropic-compat
+  endpoint as 5.1 (bare id and the `[1m]` long-context tag both answer). Rates
+  mirror 5.1 and match `PRICING`, which keeps glm-5.2 a declared placeholder.
+
 - **`numpy` was declared in no extra, so the plus contract could not run.**
   The EvalPlus expanded harnesses (`mbpp-plus`, `humaneval-plus`) `import numpy`
   inside the graded program, so under the repo's own documented

chimera/providers/catalog.py

@@ -136,6 +136,19 @@ def resolve_api_key(self) -> str | None:
                 base_url="https://api.z.ai/api/anthropic",
                 api_key_env="ANTHROPIC_AUTH_TOKEN",
                 context_window=200_000, cost=(2.0, 8.0)),
+    # glm-5.2 — the model this repo actually publishes its flagship scorecard
+    # with, and which was MISSING from the catalog until 2026-07-29: a
+    # bench-matrix run naming it failed with "Unknown provider: 'glm-5.2'" and
+    # recorded an error cell, so the model behind the published numbers could
+    # not be re-run through the CLI. Verified served by the same z.ai
+    # Anthropic-compat endpoint as 5.1 (bare id and the `[1m]` long-context
+    # tag both answer). Rates mirror glm-5.1 pending Zhipu's public sheet —
+    # PRICING/PRICING_PLACEHOLDERS in cost.py is the source of truth and keeps
+    # glm-5.2 as a declared placeholder, so this row must not diverge from it.
+    ModelConfig("glm-5.2", "anthropic",
+                base_url="https://api.z.ai/api/anthropic",
+                api_key_env="ANTHROPIC_AUTH_TOKEN",
+                context_window=200_000, cost=(2.0, 8.0)),
     # GPT-OSS (OpenAI open-weights). Distributed via Ollama; both 20B
     # and 120B run locally on adequately-sized hardware. $0/$0 because
     # local serve is free.