#2465 Follow-up to #2378: --calibrate saves values contradicting its own round winners

本文

## Follow-up to #2378: `--calibrate` completes on string-gpu duals, but saves values contradicting its own measurements

Found while verifying the #2378/#2380 fix (which works — no more crash, tuning runs to completion on `"gpu": "0,1"`). The save step writes a config that matches neither the round arrows nor the previous file. Engine `0.1.44.1` + setup `01cb90dc`, dual RTX 3090, IQ3_S 262K, `strata-iq3_s-dual-iso8.json`.

**1. PCIe share: arrow says 1, file says auto.**
Rounds (median of 3): `0: 125.8 (123.9–129.4), 0.2: 127.2, 0.35: 127.8, 0.55: 127.4, 0.75: 127.5, 0.9: 125.7, 1: 129.5 (126.4–131.7) -> best 1`. Every value overlaps; fine. But the file ends up with `--pcie-frac` **removed** (→ auto). Either the best should be written, or the verdict should say the race was tied and it kept auto — currently arrow and file disagree.

**2. Draft floor: arrow says 0.7, file says 0.5.**
`0.3: 119.8, 0.5: 129.8, 0.7: 132.1 -> best 0.7`, yet the file carries `--spec-min-p 0.5`. This one is unambiguous: the measured winner was not saved.

**3. Expert tier: raw numbers say 1-160, verdict says default.**
`default 112.5, every-1-80 111.1, every-1-160 114.9` → `[ok] default settings are already the fastest`, adapt flags stripped. The verdict contradicts its own raw table (114.9 > 112.5). If a margin/significance rule flips this, it should be printed; as-is it reads as a bug.

**Impact:** after a clean calibrate run the config matches nothing the run measured, and a user comparing arrows to file (as I did) cannot tell which to trust. Suggest either saving exactly what the arrows name, or printing the verdict rule (tie margin, cost preference) whenever it overrides a round winner.