docs/qa.md

@@ -161,12 +161,25 @@ turns look misaligned when they are not.
    **The slowness now has a cause and a number** (04-08-2026). A spoken turn
    takes 32 to 34 seconds. One phrasing call is 30.0s of that. STT is 1.0s
    and routing is under 10ms. Both interactive calls decoded exactly 512 tokens,
-   which is the `chat` cap. So both ran to the cap and were truncated, to
-   produce a 16-character reply. The resident model is a Thinking variant and nothing
-   tells it to skip reasoning. `deploy/mavgpud.json` already passes
-   `--chat-template-kwargs {"enable_thinking":false}` for the workstation and
-   `llamaArgs` does not. Filed as **531**, which also asks whether that gap
-   inflates the workstation comparison.
+   which is the phrasing cap. Both were truncated, to produce a reply of
+   under 25 characters.
+   The cause is `responseGrammar`, not the model. Its last rule is
+   `ws ::= [ \t\n]*`, and `*` is unbounded, so the model emits `{` and then
+   satisfies `ws` with whitespace until `max_tokens` stops it. Reproduced on a
+   second server: at `repeat_penalty` 1.0 it runs to 512 and returns
+   `finish_reason=length`, at 1.3 it stops at 24. Bounding the rule to
+   `[ \t\n]{0,4}` gives a clean stop at 33 tokens three times out of three with
+   no penalty at all.
+   Only some callers are exposed. `internal/llm.Req` sends `repeat_penalty` and
+   the replier sets it to 1.3, so that path is protected by accident. `chatReq`
+   in the phraser sends no penalty, so `PhraseChat`, `PhraseQuery`,
+   `PhraseNudge` and `PhraseReminder` all run at the default 1.0. Filed as
+   **531**.
+   Two guesses were wrong on the way and are recorded so nobody repeats them.
+   It is not reasoning tokens: the probe returned `reasoning_content` of length
+   0, and the grammar constrains output from the first token. It is not the
+   `--cache-ram 512` limit either: that is MiB of prompt cache and the 512 that
+   was hit is a token count.
    The wandering is a second thing and stays on 287.
 
 **The wake path cannot be checked here, and that is now the decision rather