The 1.62 arc closes. Its five chapters landed in 1.61.1 and 1.61.2, in the order the plan committed to; the minor is the release where the story finishes, and this is what finished. The properties now held — by tests that fail the build, not by sentences:
optimizenever throws, over a seeded corpus of hostile atoms: RTL and CJK text, lone surrogates, null bytes, zero-width characters, CRLF, unclosed fences, 3,000-character tokens.optimizenever grows the token count, at both levels.optimizeis idempotent — the pipeline runs to a fixed point, andoptimize(optimize(x))is byte-identical tooptimize(x).- Protected spans survive byte-for-byte — code blocks, inline code and URLs, asserted across the whole corpus, with bait the rules want to rewrite planted inside protected spans so the assertion can actually fail.
- Money is never negative — no document this package can build carries a negative dollar figure, whatever the input, and no two doors to the same value disagree about what fits through.
- Unreadable lines are named —
profileUsage,conformandrollUpnever throw on any text; what they cannot read they report by line number.
Added#
docs/hardening.md— the properties above stated as a page a stranger can read: where each is enforced, why every defect the stress session found is pinned as a named case outside the fuzzer's seed schedule, what a bounded fuzzer does not prove, and how to reproduce a verdict. Linked from the documentation index in the deciding-whether-to-trust section, next to the doctrine it borrows its standard from.
Changed#
- The documentation index's account of the plans was written when there were five and never counted past them; it now counts seven, and the arcs table says the 1.62 arc landed rather than that it is in progress.
Release mechanics#
- All five manifests, the
@trazum/coredependency in the CLI, the MCP server and the web app, and the lockfile move to 1.62.0. - The three README Action pins advance to
cd1588c, the 1.61.2 release commit.