fix(models): correct contributor guard, 1M context, docs for muse-spark-1.3

- model_data_policy_guard: name the triggering -contributor model instead
  of hardcoded 1.2; per-version verified price tables (1.3 standard
  $1.25/$4.25 via OpenRouter live metadata; cached figures 1.2-only)
- model_metadata: muse-spark-1.3 + muse-spark family at 1048576 (OpenRouter
  verified 2026-09-02) with pre-catalog stale-cache keys so 256K-fallback
  sessions self-heal
- docs: contributor-tier notes cover 1.2 + 1.3
- tests: 1.3 guard regression, muse stale-cache guard, live-catalog mirror
  gains 1.3-contributor-free (confirmed on live relay)

143 tests pass (guard, selection guards, opencode catalog, model_metadata);
ruff clean.

agent/model_metadata.py

@@ -545,6 +545,13 @@ DEFAULT_CONTEXT_LENGTHS = {
     "deepseek": 128000,
     # Meta
     "llama": 131072,
+    # Muse Spark family (1.1/1.2/1.3 + contributor tiers) ships with a 1M
+    # context window: 1,048,576 per OpenRouter live metadata (verified
+    # 2026-09-02). The family key covers every checkpoint; live endpoint /
+    # models.dev metadata still wins when available. Substring match also
+    # covers -contributor and provider-prefixed ids (meta/...).
+    "muse-spark-1.3": 1_048_576,
+    "muse-spark": 1_048_576,
     # Thinking Machines — Inkling family ships with a 1M context window
     # (max output 256K).  Verified against OpenRouter live metadata
     # (context_length 1,048,576 for inkling, inkling-small, and the
@@ -2304,6 +2311,8 @@ def _model_name_suggests_minimax_m3(model: str) -> bool:
 # catch-all can never be listed here.
 _PRE_CATALOG_STALE_KEYS = frozenset({
     "minimax-m3",    # 1M; older builds persisted the "minimax" catch-all (204,800)
+    "muse-spark-1.3",  # 1M; builds before this entry fell through to the 256K fallback
+    "muse-spark",      # 1M; 1.1/1.2 builds fell through to the 256K fallback
     "grok-4.3",      # 1M; pre-2026-05-15 builds persisted the "grok-4" catch-all (256,000)
     "grok-4.6",      # 500K; pre-catalog builds persisted the "grok-4" catch-all (256,000)
     "grok-4-fast",   # 2M; pre-2026-04-10 builds fell through to the 256K probe fallback

tests/agent/test_model_metadata.py

@@ -1631,6 +1631,18 @@ class TestGrok43StaleCacheGuard:
             assert ctx == 256_000, f"{slug} should stay 256000, got {ctx}"
 
 
+class TestMuseSparkStaleCacheGuard:
+    """Muse Spark (1M window per OpenRouter live metadata) had no catalog
+    entry, so older builds persisted the 256K default fallback. The cache
+    guard must flag that stale value and keep correct/probed values."""
+
+    def test_stale_muse_spark_detected_by_generic_guard(self):
+        from agent.model_metadata import _stale_pre_catalog_cache_entry
+        for slug in ("muse-spark-1.3", "meta/muse-spark-1.3-contributor", "muse-spark-1.2-contributor"):
+            assert _stale_pre_catalog_cache_entry(slug, 256_000), slug
+            assert not _stale_pre_catalog_cache_entry(slug, 1_048_576), slug
+
+
 class TestGrok46StaleCacheGuard:
     """Pre-catalog builds resolved grok-4.6 via the generic 'grok-4' catch-all
     (256,000) and persisted it before the 500K catalog entry existed.

tests/hermes_cli/test_opencode_free_live_catalog.py

@@ -41,6 +41,7 @@ _LIVE_FREE_MODELS = [
     "nemotron-3-ultra-free",
     "nemotron-3.5-lightning-free",
     "muse-spark-1.2-contributor-free",
+    "muse-spark-1.3-contributor-free",
 ]
 
 # The raw live /zen/v1/models dump also lists paid/subscription + KEYED-free IDs

website/docs/integrations/providers.md

@@ -329,7 +329,7 @@ model:
 Base URLs can be overridden with `NOVITA_BASE_URL`, `GLM_BASE_URL`, `KIMI_BASE_URL`, `MINIMAX_BASE_URL`, `MINIMAX_CN_BASE_URL`, `DASHSCOPE_BASE_URL`, `XIAOMI_BASE_URL`, `GMI_BASE_URL`, `META_BASE_URL`, or `TOKENHUB_BASE_URL` environment variables.
 
 :::note Meta contributor tier
-`muse-spark-1.2-contributor` is Meta's contributor tier — Meta may train on your prompts and completions, so [interactive model selection asks for confirmation](../user-guide/configuring-models.md) before using it. For current pricing and rate limits, see [Meta Model API pricing and rate limits](https://dev.meta.ai/docs/pricing-rate-limits/). Use `muse-spark-1.2` (standard variant, no training) for confidential work.
+`muse-spark-1.2-contributor` and `muse-spark-1.3-contributor` are Meta's contributor tiers — Meta may train on your prompts and completions, so [interactive model selection asks for confirmation](../user-guide/configuring-models.md) before using either. For current pricing and rate limits, see [Meta Model API pricing and rate limits](https://dev.meta.ai/docs/pricing-rate-limits/). Use the standard `muse-spark-1.2` / `muse-spark-1.3` (no training) for confidential work.
 :::
 
 :::note Z.AI Endpoint Auto-Detection

website/docs/user-guide/configuring-models.md

@@ -57,7 +57,7 @@ Prompt caches are keyed to the model serving the request, so any mid-conversatio
 
 ### Unattended data-training tiers
 
-Models such as `muse-spark-1.2-contributor` are discounted because the vendor may train on your prompts and completions. Interactive model selection always shows a confirmation prompt. Non-interactive startup paths such as Kanban workers and cron agents fail closed because they cannot ask that question.
+Models with a `-contributor` suffix (e.g. `muse-spark-1.2-contributor`, `muse-spark-1.3-contributor`) are discounted because the vendor may train on your prompts and completions. Interactive model selection always shows a confirmation prompt. Non-interactive startup paths such as Kanban workers and cron agents fail closed because they cannot ask that question.
 
 If training on the unattended workload's data is acceptable, record a persistent acknowledgement: