Consult ignore_eos when classifying a finish reason at the output limit

A request with ignore_eos suppresses end-of-sequence termination, so its last
sampled token can be an EOS id while it finishes on the output limit. The
terminal event classified that as a stop with Eos termination, which tells
clients a truncated generation completed normally.

Use the request's ignore_eos setting in the finish-reason classification, the
same way the emitting and logprob paths already do, and document the finish
reason the client sees for each termination cause.

Validation: release Helios library suite on the build host, 1118 passed and 5
failed; the same tree without this change, with the new regression case
applied, is 1117 passed and 6 failed, and the five shared failures are the
pre-existing engram staging and package-cache environment failures reproduced
on the unmodified tree. Both builds also served one prepared single-GPU package
side by side: with greedy sampling, a request that reaches its output limit on
an EOS token reports stop on the unmodified build and length on this one, 6/6
each way, while requests that honour EOS, stop before it, or ignore it without
ending on it report the same reason on both builds. Material limits: the
regression drives the simulated worker phases, and the sampled-EOS suppression
itself is unchanged.

Design-Signed-Off-By: Hermes Agent (DeepSeek V4.1 Flash); design-review=2d93

Author: mark-yong

docs/TODO.txt

@@ -98,9 +98,6 @@
 [ ] Qualify prompt scoring on remaining recipes and cache lifecycles;
     see docs/logprobs-design.md for TP4 costs and numerical limits.
 
-[ ] Respect ignore_eos when classifying EOS at the output limit;
-    see helios/src/server/runtime.rs.
-
 [ ] Validate embedding readiness timestamps in full Qwen3.8 PLE serving and
     cross-node RoCEnante TP packages for Qwen3.8 and DeepSeek V4.1.
 

docs/api.md

@@ -352,6 +352,13 @@ or partial tags. Structured answers become verbatim when ordinary answer content
 begins, preserving scalar JSON, Lark output, and markup inside string values.
 Shared runtime handling owns token accounting, stop filtering, and termination.
 
+Termination supplies the client finish reason: `length` at the output limit,
+`tool_calls` when parsed calls complete an answer that the user did not stop, and
+`stop` for a matched token or string stop or a terminating end-of-sequence token.
+A request admitted with `ignore_eos` does not treat end-of-sequence tokens as
+terminating, so one that exhausts its output limit reports `length` even when its
+last generated token is end-of-sequence.
+
 A reasoning delimiter that the tokenizer encodes as one token, such as
 `<think>`, `</think>`, `</mm:think>`, or Gemma 4's `<channel|>`, opens or
 closes reasoning only when the model generates that token. When a model writes

helios/src/server/runtime.rs

@@ -1594,15 +1594,11 @@ impl<'a> Worker<'a> {
                                 );
                             }
                         }
-                        let reason = if user_stop
-                            || request
+                        let eos = !request.incoming.prompt.ignore_eos
+                            && request
                                 .last_token
-                                .is_some_and(|token| self.config.eos_token_ids.contains(&token))
-                        {
-                            "stop"
-                        } else {
-                            "length"
-                        };
+                                .is_some_and(|token| self.config.eos_token_ids.contains(&token));
+                        let reason = if user_stop || eos { "stop" } else { "length" };
                         if request
                             .incoming
                             .events