fix(common): tolerate dangling Gemma 4 channel prefix after tool responses

[rodboev]

Overview

Gemma 4 tool-calling turns can end on a bare <|channel> prefix with optional trailing whitespace. On current master, that fragment is treated as invalid peg-gemma4 output, so llama-server returns a 500 with:

The model produced output that does not match the expected peg-gemma4 format

This change keeps the fix in the shared Gemma 4 parse path. It tolerates only the exact final dangling <|channel> prefix with trailing whitespace, preserves any already parsed assistant state, keeps the existing tolerated trailing marker cases, and still rejects malformed non-whitespace tails after parsed prefixes.

Closes #25072

Additional information

The live issue log shows the final unparsed fragment as:

<|channel> 

The non-tool Gemma 4 path already has narrow tolerance for trailing channel-marker noise. The tool-enabled path did not have an equivalent final-parse escape hatch, which is why the shared parser escalated this exact fragment into a 500.

Focused validation for the implemented diff stayed on the parser surfaces:

test-chat.exe --template google-gemma-4-31B-it.jinja --detailed
ctest --test-dir build-target1-msvc-nofatal -R "^test-chat$" --output-on-failure --timeout 900
ctest --test-dir build-target1-msvc-nofatal -R "^test-chat-peg-parser$" --output-on-failure --timeout 900

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES - AI assisted with local issue triage, patch planning, and test planning.

[aldehir]

Thanks for the PR, but we do not indiscriminately apply fixes to code that impacts all models. Gemma 4 has a dedicated parser which should encode this edge case, iff it exists with a well-behaved client.

First and foremost, due diligence should be done to ensure the reporter is using a proper client that sends back reasoning content. The absence of reasoning content from tool calls leads to model hallucinations and is the source of many parsing problems.

[rodboev]

That makes sense. I see you already asked, so we we should have more details about the tool responses soon. In the meantime I'm moving the fix into the dedicated Gemma 4 grammar rather than the shared parser path.

[magicmaaaaan]

This is probably from a bug in the chat template, and it should be noted that there's a slew of upcoming fixes in https://huggingface.co/google/gemma-4-31B-it/discussions/118. I believe "fix: restore model turn + thinking cue after tool responses" is the relevant fix to this particular issue.