What's new
Non-English session continuity (#127, #128) — the topic tokenizer now handles Korean, Chinese, Japanese, and accented Latin across all runtimes (Claude, Codex, OpenClaw, OpenCode). Korean prompts (including common 2-char nouns) and accented Spanish/French now match checkpoints instead of scoring a hard 0.0. Reported by @yujung7768903.
Dashboard hero (#130) — the "plan goes further" headline leads with the per-token throughput % again (e.g. "31.3% more work per token") instead of the dollar-slice string; the $ figures stay in the per-window cards.
Fixes (#131) — greened the 4 pre-existing Windows-only test failures (HOME vs USERPROFILE home resolution; POSIX file-mode assertion), plus three pre-existing continuity bugs surfaced by review (JS \z regex, missing "continue from" cue, strip-first-vs-strip-all parity).
Full parity verified across Python, OpenClaw, and OpenCode via a shared test fixture.
