fix(realtime): accept bounded partial VAD overrun

[LauraGPT]

Summary

• accept a proven long-segment partial when its decode endpoint is slightly later than the VAD endpoint
• bound the allowed overrun to one realtime decode chunk
• add exact regression coverage for the reporter's 114430/126976/126740ms trace and a farther-overrun rejection test

Root cause

The follow-up on #3302 used exact merge 2b6294d69e6588a63c0ce06300f1b55ded71ae58 and still locked a complete partial down to 哎. The last partial covered 114430-126976ms, while FSMN-VAD emitted 114430-126740ms.

The long-segment reconciliation gate only accepted tail_gap_ms >= 0. Normal decode/VAD scheduling therefore rejected this candidate solely because it ended 236ms after the VAD boundary, before the single-character regression check could run.

The new lower bound allows at most one decode chunk of overrun. Existing exact start alignment, repeated observations, long-segment duration, text alignment, and hallucination gates remain unchanged. A partial more than two chunks past this concrete boundary remains rejected by regression coverage.

Verification

Exact signed+DCO head: 980765099dea8d9f34ea2f10889b3ab1c3876173

• before the fix, the exact reporter-offset test returned 哎
• focused boundary tests: 2 passed
• realtime/postprocess/file-finalization/docs suites: 141 passed, 1 existing environment skip
• Python compilation and git diff --check: passed

Related to #3302. Merging this PR is not reporter confirmation; keep the issue open until the original file-upload and microphone paths are retested.

Signed-off-by: LauraGPT 18321252+LauraGPT@users.noreply.github.com