docs(readme): measured comparison of the judging backends

[ChenYCL]

What

One new README section — ## Judging backends: measured comparison in README.md, mirrored as
## 判定后端实测对比 in README.zh-CN.md — carrying the measured comparison of the judging
backends that until now lived only in experiments/** and docs/**.

It contains: the 20-item accuracy table (hosted Jev, Kev 4B at both row limits, the GGUF 4B
local-readout default, the two other 4B candidates, Kev 0.8B, the GGUF 0.8B and the trivial
positional baseline); what each costs in per-item latency, disk, memory and money; a short "how this
was measured"; the 15 real browser runs; and the three goal_done bars with their measured basis.

No code, tests or config changed. The existing "hosted Jev is the default" statement is untouched —
the new section restates it. The only edit outside the new section is lives in SKILL.md →
also lives in SKILL.md (and 见 → 同时见), so the pointer above the new tables no longer reads as
if the numbers were only in SKILL.md.

Sources for every figure

figure	source
hosted Jev 19/20 = 0.95 · browser 14/15 · noul 5/5 · action+click 5/5	experiments/gguf-provider/results/ceiling-jev-20items.json (overall, by_kind); restated in experiments/gguf-provider/RESULTS.md and results/local-models-4b.md §0
hosted Jev 621 / 1,270 ms per item, $0.0035 for the 20 items	results/ceiling-jev-20items.json (latency_ms, cost_usd_total)
Kev 4B @16384: 19/20 · 14/15 · 5/5 · 5/5 · 2,223 / 12,183 ms · 36 GB footprint · 23,769 ms packed step	experiments/kev-4b/README.md (Measured: Kev 4B / Memory); cross-checked in results/local-models-4b.md §0
Kev 4B @8192: 18/20 · 13/15 · 5/5 · 4/5 · 1,965 / 8,995 ms · 12.2 GiB RSS peak	same two files
Kev 4B disk 9.34 GB base + 152 MiB checkpoint; Kev 0.8B 1.72 GiB download	docs/local-kev-bringup.md §0, experiments/kev-4b/README.md (verification log)
Kev 0.8B 14/20 (14 of 19 answerable) · 9/15 · 5/5 · 3/5 · 590 / 6,353 ms · 3.6 GiB peak	experiments/kev-4b/README.md (Measured: Kev 0.8B, Row limit §3)
GGUF 4B 16/20 · 12/15 · 4/5 · 2/5 · 3,072 / 13,895 ms · 3,362 MiB RSS · 2,740,937,888 B	experiments/gguf-provider/RESULTS.md, results/local-models-4b.md §0/§2
Qwen3-4B-Instruct-2507 15/20 · 10/15 · 5/5 · 1/5 · 2,514 / 15,386 ms · 2,497,281,120 B	results/local-models-4b.md §0/§2
gemma-3-4b-it 10/20 · 7/15 · 3/5 · 1/5 · 2,189 / 9,725 ms · 2,489,894,016 B	results/local-models-4b.md §0/§2
GGUF 0.8B 10/20 · 7/15 · 3/5 · 1/5 · 777 / 4,178 ms · 811,843,840 B	experiments/gguf-provider/RESULTS.md
baseline 11/20 · 9/15 · 2/5 · 2/5	results/baseline-first-option.json + RESULTS.md
clean GGUF 4B step 18.3 s cold / 9.0 s warm, 6.1 s of it click_target	results/local-models-4b.md §0.2, RESULTS.md
rotation k=0/3/7 → e8/e5/e1, P≈0.99	experiments/gguf-provider/RESULTS.md
15 runs: 4 success / 8 stuck / 2 needs_user / 1 max_steps; 39 step requests; 0 timeouts; 0 LOW_LABEL_MASS; min label mass 0.870	docs/local-backend-run-smoke.md §0/§2/§3
click_target 0.71–0.78 on a 65-element page; type last of four (0.105 vs 0.430); 3 premature stops; 8 of 9 non-success runs stop on action	docs/local-backend-run-smoke.md §0/§5
bars 0.174 / 0.482; bands 0.12–0.28 and (0.341, 0.683]; 7/0/0 vs 4/0/3	docs/local-backend-run-smoke.md §4/§9, experiments/kev-4b/README.md (Threshold replay)

Nothing is estimated. Cells the records do not measure say not measured; cells that do not exist
for a backend say —.

Verification

• Both new tables parse as consistent GFM grids (equal cell counts per row), code fences are
  balanced, and every relative link in the new section resolves to a file that exists.
• CI: npm test (mock mode, Ubuntu + headless Chrome) — docs-only change.