#2603 docs: DETAILS.md: since 0.1.42 greedy repeats also need a fixed --pcie-frac (the share is refined at the start, #1928)

本文

The docs part of #1928.

**What it changes.** Two places in `docs/DETAILS.md`:
- The "Reproducible greedy output" paragraph named `STRATA_IQ_MT_MIN=1` and, for byte-identical repeats, `--pcie-frac 0`. It now adds that since 0.1.42 the PCIe share of the missed experts is refined from measured times in the first decode windows (one GPU, CUDA or RDNA4, no `--pcie-frac`; linked to the section that describes it) and can end differently on two fresh servers, with #1928's numbers: one reporter's RTX 5090 (Windows, serve mode, IQ3_S, 0.1.42, 5 fresh servers per prompt, temperature 0, 256 tokens): `STRATA_IQ_MT_MIN=1` alone 1, 1 and 4 different answers on three prompts; with a fixed `--pcie-frac 0.55` as well 1, 1 and 1 (10 servers on the third). A fixed `--pcie-frac` stops the refinement; `STRATA_PCIE_FRAC_DEFAULT=old` should too and is marked as not tried.
- The sampling paragraph's "for seed-reproducible output add `--adapt-every 100000`" adds "and since 0.1.42 a fixed `--pcie-frac` as well (#1928, measured at temperature 0)".

**What it leaves alone.** The code, including the comment in `include/strata/spec/draft_policy.hpp` that #1928 also names (an engine header; a code-comment change needs the backends built, so it is left for an engine PR), and every other line.

**How it was checked.** Against #1928's tables and DETAILS.md's PcieAuto section (scope, "`--pcie-frac N` always wins"); the anchor matches GitHub's slug of that heading. A sub-agent reviewer checked it; its corrections ("no longer always", the reporter's setup, the untested opt-out, the softer sampling clause) are in.

🤖 Generated with [Claude Code](https://claude.com/claude-code)