fix: preserve completion usage and trim stop tokens - #312

Description

Preserve OpenAI prompt-token usage across sequence re-entry and trim model stop tokens consistently from completion text.

Motivation

Sequence re-entry mutates the internal prompt length while recomputing KV state, but response usage must retain the caller's original prompt. Two completion detokenization paths also bypassed the existing EOS-aware helper and could leak stop tokens.

Milestone (optional)

GLM-5.2 release closeout.

Type of Change

• model — add/extend model support (models/** + registration seam + model kernels only)
• kernel — add/optimize a compute kernel (batchgen_kernels/** + in-tree kernel dirs)
• core — change scheduling/serving/runtime scaffolding (the only type that may)
• fix — narrow bug fix (+ a regression test)
• infra — build / CI / packaging / scripts / Docker
• docs — documentation only

File changes

Validation

• PR hygiene: PASS for PR_TYPE=fix.
• Focused remote tests passed on the aggregate release branch containing the identical patch.
• Clean reconstructed-branch remote gate: pending before ready-for-review.

Checklist

• I have read the CONTRIBUTING guide and PR Merge Policy Contract.
• I have updated the tests.
• I have updated the documentation if applicable.

PR Merge Policy Contract — pre-merge checklist

• Diff reviewed; every file traces to the task.
• Exactly one type is selected and all files fit it.
• File table matches the diff.
• No scratch/debug production scripts.
• Tests are under tests/.
• No new environment-variable debug guard.
• No leftover debug output or commented-out code.
• No artifacts staged.
• No public API change requiring module docs.
• One surgical concern.
• Commit format and body comply with policy.
• CI is green.