Fix: carry_initial_prompt overflows the decoder prompt budget

[Yigtwxx]

Problem

carry_initial_prompt=True exists so that a long initial_prompt (a glossary, a
list of proper nouns) keeps reaching the decoder on every sliding window. When
the prompt is long enough to fill the decoder's prompt budget, it does the
opposite: from the second window onwards the initial_prompt is dropped
entirely and replaced by carried-over transcript.

Root cause

The prompt budget is n_text_ctx // 2 - 1, which is 223 tokens for the released
checkpoints (n_text_ctx = 448). remaining_prompt_length tracks how much of
that budget is left once initial_prompt has taken its share:

remaining_prompt_length = model.dims.n_text_ctx // 2 - 1
if initial_prompt is not None:
    initial_prompt_tokens = tokenizer.encode(" " + initial_prompt.strip())
    all_tokens.extend(initial_prompt_tokens)
    remaining_prompt_length -= len(initial_prompt_tokens)   # may reach 0 or go negative

if carry_initial_prompt:
    nignored = max(len(initial_prompt_tokens), prompt_reset_since)
    remaining_prompt = all_tokens[nignored:][-remaining_prompt_length:]
    decode_options["prompt"] = initial_prompt_tokens + remaining_prompt

Once initial_prompt is 223 tokens or longer, remaining_prompt_length is zero
or negative and the slice inverts:

>>> tokens = [1, 2, 3, 4, 5]
>>> tokens[-2:]     # room for 2 more tokens
[4, 5]
>>> tokens[-0:]     # no room left - returns everything
[1, 2, 3, 4, 5]
>>> tokens[-(-2):]  # negative - drops from the front instead
[3, 4, 5]

So instead of carrying nothing, the window carries the entire accumulated
transcript. DecodingTask._get_initial_tokens then trims the combined prompt
from the left:

prompt_tokens[-(self.n_ctx // 2 - 1) :]

Left-trimming removes the leading tokens, which are exactly the
initial_prompt tokens the option is meant to preserve.

Fix

Clamp the remaining length at zero, and slice explicitly so the empty case
stays empty.

Evidence

tiny.en on CPU, tests/jfk.flac tiled to ~66 s (3 windows), temperature=0
and condition_on_previous_text=True so the carry-over path runs on every
window. DecodingTask.__init__ was wrapped to record options.prompt.

initial_prompt of exactly 223 tokens (budget exactly full, so nothing may
be carried):

window	prompt length	tokens over budget	initial_prompt reaches the decoder
0	223 → 223	0 → 0	yes → yes
1	284 → 223	61 → 0	no → yes
2	367 → 223	144 → 0	no → yes

initial_prompt of 14 tokens (room to carry, i.e. every existing use of the
option):

window	prompt length before	prompt length after
0	14	14
1	76	76
2	134	134

Per-window prompt token ids are identical before and after, and the transcript
is byte-identical (sha256 861f8b56223f5e5124744caacf57f1f87dc4b37a10472cecf9a5ab6abf355bc5
in both runs).

The repository's own workflow was run on this branch before opening the PR and
is green across the whole matrix, pre-commit included:
https://github.com/Yigtwxx/whisper/actions/runs/30698478749

Compatibility

No API, CLI, or default behavior change. The only inputs affected are those
where remaining_prompt_length was already zero or negative, i.e. an
initial_prompt of at least n_text_ctx // 2 - 1 tokens combined with
carry_initial_prompt=True. Every shorter prompt takes the same code path as
before and produces identical output.