Skip to content

Release 0.12.0 - #173

Merged
silverstein merged 1 commit into
masterfrom
release/0.12.0
Aug 21, 2026
Merged

silverstein merged 1 commit into
masterfrom
release/0.12.0

Conversation

@silverstein

Copy link
Copy Markdown
Member

Cuts 0.12.0 from the twenty commits on master since v0.11.0.

Why minor rather than the planned 0.11.1. convert_pdf_to_markdown now writes lowered smaller glyph runs as Unicode subscripts, so the same PDF produces different Markdown: 672 characters across 48 of 55 pages of the Shannon paper, every one a pure substitution. That is a user-visible output change, not a bug fix, and anyone diffing extracted Markdown will see it. This repo has spent the last two days fixing version strings that lied about what they contained, so calling this a patch would be the same mistake in a smaller font.

Also in it. The plugin publish guard is bound to a build it can identify (#168), the freshness gate's shipped-path list is bound to what the build actually copies, the gate gets the repository instead of a depth-1 snapshot of its tip (#169), a second writer to the native canvas marker is reported as a second writer rather than as a resource limit, and the load-sensitive Docling helper suite got a timeout it chose for itself instead of inheriting one.

Eight version sites, not seven. The root package-lock.json still said 0.10.0. It was missed for 0.11.0 as well, so it had drifted through two releases without anything failing. It is not cosmetic: the layout occurrence oracle binds package-lock.json by digest, and npm rewrites that field on the next npm install, so a stale root lock turns an unrelated install into a silent oracle invalidation that surfaces as scorer failures with no visible cause. Bumped here and added to docs/RELEASE.md with the reasoning, alongside the four sites that were missing before 0.11.0.

Fixtures were regenerated by their own generators rather than by hand: the layout occurrence oracle, which binds package.json and the lock, and the pinned tool contract, which carries the advertised version (44 tools).

Verified: smoke-mcpb-contract 9/9 (the test the 0.11.0 bump broke by moving only three of the sites), mcp-contract and extraction-phase1-layout-evidence 52/52, trajectory-grader version-pinned tool contract passing. The full suite is CI's to run.

🤖 Generated with Claude Code

Minor, not a patch. convert_pdf_to_markdown now writes lowered smaller glyph
runs as Unicode subscripts, so the same PDF yields different Markdown: 672
characters across 48 of 55 pages of the Shannon paper. Anyone diffing output
sees it, and calling that a bug fix would understate it.

Also carries the release-infrastructure work: the plugin publish guard is bound
to a build it can identify, the freshness gate's shipped-path list is bound to
what the build actually copies, the gate gets the repository rather than a
depth-1 snapshot of its tip, a second writer is reported as a second writer
rather than as a resource limit, and one load-sensitive suite got a timeout it
chose for itself.

Eight version sites, not seven. The root package-lock.json still said 0.10.0,
having been missed for 0.11.0 too. Nothing fails when it drifts, which is why
nobody noticed, but the layout occurrence oracle binds package-lock.json by
digest and npm rewrites that field on the next install, so a stale root lock
turns an unrelated `npm install` into a silent oracle invalidation. Added to
the checklist with that reasoning.

Fixtures regenerated by their own generators: the layout occurrence oracle
(binds package.json and the lock) and the pinned tool contract (carries the
advertised version, 44 tools).

smoke-mcpb-contract 9/9, mcp-contract and layout-evidence 52/52,
trajectory-grader tool contract passing.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@silverstein
silverstein merged commit 0baf06b into master Aug 21, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant