docs(onnx-pack): where the numbers live points at the tree, not a private state file - #685

VERDICT: READY (tprs-ready applied). The private-path pointer is gone, both replacement links resolve, and the one number in the new paragraph is supported by the page it points at. Two small precision notes.

• No private path left in docs/onnx-pack.md at this head (grep -nE "~/|/home/|briefs|TORTEP" → nothing). That was the whole point: a public page was telling a reader to look in ~/briefs/TORTEP-STATE.md.
• Both links resolve: ../onnx/BENCHMARK-PARITY.md → onnx/BENCHMARK-PARITY.md (present on main since #683, 201 lines: the AMD protocol, the two --cached CPU tables, and the ## NPU row: our engine, the dx route section "driven by onnx/bench_npu.sh" — which exists), and onnx.md → docs/onnx.md.
• "within 0.5 %" is supported by the page it links. docs/onnx.md's Status section records the device proof as max_abs = 9.893e-02 against max|ref| = 20.414, i.e. rel 4.846e-03 against a 2e-2 gate — 0.48 %, which is the 0.5 % claimed. The two are consistent; they are just in different forms.

Note 1 (precision): "the CPU rows (ryzen and Strix Halo, five prompt lengths, default and --cached exports)". BENCHMARK-PARITY's tables are --cached rows only — the default/uncached comparison lives in onnx/README.md (uncached 291.22 ms/token vs --cached 63.66, 4.6x at 256 tokens). Either say "the --cached rows" here, or name both files.

Note 2 (precision, so the number does not drift): the page states the tolerance as rel 4.846e-03 and the paragraph states "0.5 %". Both are right; anyone re-deriving will want the pair tied together, e.g. "within 0.5 % (rel 4.8e-03 on the reference shape)". Not a blocker.

No numbers were changed or added beyond that one, the diff is a single paragraph (+6/−3), and the stale "No KV cache" the owner saw on the live page is a site build-schedule artefact, as the body says — 1bit.gg was last built at 05:36Z, before #684 merged.