[leehack]

Docs only. This records the final #174 re-gate at bee747a5 and the maintainer's acceptance of the two performance costs below 0.97×. It also marks native v0.9.0 as published.

Native v0.9.0 was published 2026-11-08 by run 42433633399 from a604639. The tag commit is a604639. The release is not a pre-release, and its SHA256SUMS lists all 14 bundles.

docs/native174_f16_dot_accumulation.md, "Before publication":

• The F16 test passes on:
  • macOS arm64;
  • all 7 Android variants on 5 devices (24 variant/device pairs; the remaining pairs are skipped because those CPUs can't run that variant);
  • GCE linux_armv8.2_2 and linux_armv8.6_1 (from the Make patch 0005 near upstream speed with in-line F16 checks (#174) #182 audit).
• v0.8.0-3 and the unpatched build fail on macOS and on Android armv8.2_2/armv8.6_1.
• Throughput:
  • S25 armv8.2_2 Q8_0 FA-off pp512: 0.77× → 0.97× of v0.8.0-3.
  • GCE Linux: ≥ 0.97× of unpatched everywhere.
  • The other Android rows below 0.97× of both baselines are listed as measured.
• Qwen3-TTS cancel: all 36 samples are within the 500 ms budget.
• Accepted costs:
  • macOS M4 F16 pp512: 0.92–0.95× of v0.8.0-3, 0.90–0.95× of unpatched. This comes from 0006; without it the build fails 21 checks.
  • S25 android_armv8.6_1 Q8_0 FA-off pp512: 0.96× of unpatched. This comes from 0005's F16 KV dot in prefill.

docs/v090_qualification.md: the "not published" paragraph now records the publication.

Release notes are generated from provenance fields only (scripts/release_publication.py), and publication checks them for an exact match. They therefore do not describe the F16 fix or the TTS additions.