#1813 Add exact speculative-window token parity and coverage gate

说明

Adds an opt-in standard-library gate comparing full token IDs and finish reasons across draft floors, serial/pipeline windows, forced rollback and repeated references. It fails on divergence, prompt reuse, incomplete requests, ignored switches or missing required coverage. It changes no engine defaults.

Refreshed to main `61b3fb5d`: all 48 pipeline comparisons and 12 default-dispatch comparisons pass with exact routing controlled by `STRATA_ROUTE_TAIL_SKIP=0`. The separate 24-request matched local/TCP comparison passes on current PR #1748 (`f2673bd2`), which supplies TCP support absent from plain main. Fifteen GPU-free gate tests pass. An explicit `matched-history` mode compares adaptive arms at the same repeat and prompt while rejecting reordered histories; the default still requires stable repeated references.

The eight-request suffix control matches tokens but correctly fails because pipeline suffix coverage is zero. Earlier default rank-7 routing differences reflect an intentional approximation whose expert decisions depend on all rows in a window; they do not establish a rollback defect. The docs now specify tail skipping, CPU arithmetic and placement controls separately.

[Fresh manifests, full outputs, engine logs and limits](https://github.com/CC-David-CC/Strata-a5500/blob/cd4d0670/bench/results/2026-10-11-spec-window-refresh/README.md). Historical evidence remains separately dated. These finite P4 checks do not certify dynamic adaptation, sampled decoding, arbitrary pipeline depth, other backends or remote overlapped decode.