docs(usage): note Ollama batch-token semantics + Prometheus/billing query divergence

Round-8 claude-review (no blockers; comment-only):

- 🟡 Documented that Ollama's /api/embed prompt_eval_count is assumed
  batch-level total and is unverified against a live instance (Ollama isn't the
  Cloud billing provider); if it proves last-item-only, switch to per-item
  summing. The char estimate already covers versions that omit the field.
- 🟡 Noted on the astrolabe_embedding_tokens_total counter that operation="query"
  is recorded pre-Qdrant, so it can legitimately exceed the billing-store
  tokens_embedded aggregate when a search fails post-embed — dashboards
  shouldn't alert on that healthy gap.

Deferred (reviewer: "minor nit, acceptable"): record_indexing_usage awaited in
the task group — the group awaits all child tasks regardless, the write is
best-effort + fast, and start_soon would need the tg threaded into the closure
for marginal gain.

Deck #284.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

nextcloud_mcp_server/observability/metrics.py

@@ -342,6 +342,12 @@ embedding_chars_total = Counter(
 # measure, Deck #67). On a dedicated counter (not folded into the chunk/request
 # metrics above) so query embeds don't inflate indexing dashboards; labelled by
 # operation = index | query. Always emitted, independent of USAGE_METERING_ENABLED.
+#
+# Dashboard note: operation="query" is recorded at embed time (before Qdrant /
+# verify-on-read), whereas the billing-store tokens_embedded row is written only
+# after the search fully succeeds. So this counter can legitimately exceed the
+# billing aggregate when a search fails post-embed — don't alert on that gap as
+# a divergence bug.
 embedding_tokens_total = Counter(
     "astrolabe_embedding_tokens_total",
     "Total embedding tokens consumed (provider-reported or estimated)",

nextcloud_mcp_server/providers/ollama.py

@@ -164,6 +164,13 @@ class OllamaProvider(Provider):
             if self._dimension is None and data["embeddings"]:
                 self._dimension = len(data["embeddings"][0])
 
+            # ``prompt_eval_count`` is assumed to be the batch-level total for a
+            # multi-input /api/embed call. Ollama's API docs aren't explicit
+            # about batch aggregation; if a version reports only the last
+            # input's tokens this understates the batch. Unverified against a
+            # live instance — Ollama isn't the Cloud billing provider (Mistral
+            # is). If it proves last-item-only, switch to per-item requests and
+            # sum. The char-based estimate covers versions that omit the field.
             prompt_eval = data.get("prompt_eval_count")
             total_tokens += (
                 round(prompt_eval)