Apply full-context fix to timestamped decoding too

The half-context cap removed for no-timestamp decoding was never
actually tied to timestamps - it originated as half of n_text_ctx in
OpenAI's reference implementation and applied regardless of the
<|notimestamps|> prompt token. Timestamped decoding now also uses the
full remaining decoder context (total_max_length - start_step).

Verified on large-v3 with the Armenian/Georgian samples from the
previous commit using timestamped prompts: Armenian now completes
fully (304 tokens vs. the previous 224-token cutoff), and Georgian
reaches the full ~445-token decoder budget instead of stopping at 224.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VhE79Fs7AknnXa9EmagCuY

Author: Emre-Akgul

python/tests/test_transformers.py

@@ -865,7 +865,7 @@ def _get_features(audio):
             assert transcription == expected_transcription
 
     @test_utils.only_on_linux
-    def test_transformers_whisper_no_timestamps_full_context(self, tmp_dir):
+    def test_transformers_whisper_full_context(self, tmp_dir):
         import transformers
 
         model_name = "openai/whisper-tiny"
@@ -907,11 +907,14 @@ def _generated_length(prompt):
         ]
         timestamped_prompt = ["<|startoftranscript|>", "<|en|>", "<|transcribe|>"]
 
-        # With the default max_length of 448, no-timestamp decoding can use up
-        # to 445 positions (448 minus the 3-token prompt prefix), exceeding the
-        # previous 224-token cap. Timestamped decoding keeps that 224 limit.
+        # With the default max_length of 448, decoding can use the entire
+        # remaining decoder context (448 minus the prompt length), no longer
+        # capped to half of max_length. This applies to both no-timestamp and
+        # timestamped decoding, since the previous 224-token cap was not
+        # actually related to timestamps. The timestamped prompt is one
+        # token shorter, so it gets one extra position of context.
         assert _generated_length(no_timestamps_prompt) == 445
-        assert _generated_length(timestamped_prompt) == 224
+        assert _generated_length(timestamped_prompt) == 446
 
     @test_utils.only_on_linux
     @test_utils.on_available_devices

src/models/whisper.cc

@@ -301,10 +301,7 @@ namespace ctranslate2 {
       decoding_options.length_penalty = options.length_penalty;
       decoding_options.repetition_penalty = options.repetition_penalty;
       decoding_options.no_repeat_ngram_size = options.no_repeat_ngram_size;
-      decoding_options.max_length = without_timestamps
-                                      ? total_max_length - start_step
-                                      : std::min(total_max_length / 2,
-                                                 total_max_length - start_step);
+      decoding_options.max_length = total_max_length - start_step;
       decoding_options.sampling_topk = options.sampling_topk;
       decoding_options.sampling_temperature = options.sampling_temperature;
       decoding_options.num_hypotheses = options.num_hypotheses;