fix(llm): correct deepseek-v4-flash pricing and the build_merge docstring

Update the deepseek-v4-flash per-1M-token pricing (same units/ordering as neighboring
models) and fix the build_merge docstring, which claimed it saved the graph back to disk;
build_merge only reads the existing graph and returns the merged result — callers persist.

graphify/build.py

@@ -1633,7 +1633,11 @@ def build_merge(
     dedup_llm_backend: str | None = None,
     root: str | Path | None = None,
 ) -> nx.Graph:
-    """Load existing graph.json, merge new chunks into it, and save back.
+    """Load existing graph.json and return it merged with ``new_chunks``.
+
+    Does NOT write to disk — the caller persists the result, e.g. via
+    ``export.to_json(G, communities, graph_path, force=True)`` after
+    clustering. ``graph_path`` is read-only here.
 
     Re-extracted files REPLACE their prior contribution per tier (#2333/#2336):
     a source_file present in new_chunks has its existing nodes/edges dropped

graphify/llm.py

@@ -172,7 +172,10 @@ BACKENDS: dict[str, dict] = {
         "default_model": "deepseek-v4-flash",
         "env_key": "DEEPSEEK_API_KEY",
         "model_env_key": "GRAPHIFY_DEEPSEEK_MODEL",
-        "pricing": {"input": 0.14, "output": 0.28},  # USD per 1M tokens (v4-flash)
+        "pricing": {"input": 0.44, "output": 1.32},  # USD per 1M tokens (v4-flash,
+        # peak, cache miss). Peak is 01:00-04:00 and 06:00-10:00 UTC Mon-Fri;
+        # all other hours are off-peak at half these rates. A cache hit is
+        # $0.014/1M in. Source: api-docs.deepseek.com/quick_start/pricing
         # deepseek-reasoner silently ignores temperature; deepseek-chat / v4-flash
         # accept 0-2, so sending 0 is safe. Note: deepseek-v4-flash (and v4-pro) have
         # thinking ENABLED by default (verified against the live API, #1621) — set