Fix ten findings from the v0.1.1 release review

[Deep-Sixed]

Fixes the ten findings from a review of what shipped in v0.1.1 (fb8af1e..5836a79). Each finding was reproduced before it was fixed. Eight commits, each self-contained with its own regression tests; reviewing commit-by-commit is the easiest path.

#	Finding	Commit
1, 2	Publish workflow uploaded an untested rebuild; the job holding the PyPI token ran mutable action refs	a141d08
3	reject(replacement=…) could half-write, or rewrite an existing record; assume() silently dropped evidence_ids	327bbf7
4, 6	Recheck swallowed an evidence-contract violation as a task failure; an auditor outage permanently failed finished work, silently	304321c
5	0.1.1 reported __version__ 0.1.0	3dad969
7	Explicit-id decisions could make a correct graph unloadable	8f6dda1
8	GRAPH.md installed loose at site-packages/GRAPH.md	0b684b1
9	Oversized LLM responses retried as transport failures	c089038
10	Evidence text stripped before its size bound	97605bc

Decisions a reviewer should weigh

Recheck semantics change (4, 6). #29/#33 made a recheck auditor exception move a DONE task to FAILED. Nothing re-runs a FAILED task, so a transient provider outage permanently failed verified work, and mesh_recheck reported only audited: 0. Now:

• The task stays DONE.
• The error is recorded as a new AUDIT_ERROR ledger event and reported in a new audit_errors field of the mesh_recheck payload.
• The next recheck simply asks again.

An auditor that raises during the initial run still fails its task, because that work was never verified. The #29/#33 tests are updated to the new contract. An EvidenceRecheckError raised inside ctx.disproved() is now fatal, the same as when it's detected after the auditor returns.

Explicit-id decisions are fixed after decide() (7). GRAPH.md checks the digest against the node's actual edges, "not against whatever its attrs claim", so a fingerprinted edge added later genuinely changes the decision. add_edge now refuses one on an explicit-id decision and points to decide(). Unaffected:

• depends_on into another decision (the Runner's task ordering)
• auto-minted decisions
• repeating an existing edge

as_of_graph rewinds the digest to the projected edges rather than special-casing the loader, per its own "not a special case" rule. GRAPH.md gains one paragraph describing this. The repo's own test_edge_binding fixture built exactly this pattern, and on 5836a79 its graph fails to load with SnapshotError. The fixture now passes the dependency to decide().

The schema moved (8). GRAPH.md now lives at contextmesh/GRAPH.md. ONTOLOGY_FILE, package-data and the README link follow it. The root GRAPH.md is now a short pointer, so external links still land somewhere; nothing reads it.

Workflow pins stay on current majors (2). Every action is pinned to a full SHA at the latest release within the major already in use. pypa/gh-action-pypi-publish v1.14.2 is dc37677, exactly where release/v1 points today, so nothing it runs changes. Upstream has newer majors (checkout v7, artifact actions v7/v8); upgrading to those is left as a separate change. There's no Dependabot config, so the pins won't refresh themselves.

The publish workflow (1)

The artifact is built once. publish now needs three jobs:

• build refuses a tag whose commit isn't on main, and pins build==1.6.1 and twine==7.0.0.
• test runs the full suite at the tagged commit.
• verify, on Linux and Windows, uses no checkout, so the source tree can't mask a packaging mistake. It installs the exact uploaded wheel and checks:
  • __version__ matches the tag
  • the schema is inside the package, with nothing loose
  • the core install has no MCP SDK
  • the CLI runs, and a session survives across two processes
  • the [mcp] extra's launcher works
  • the sdist builds and installs the same package

publish then uploads those same files.

Verification

• 1009 tests pass (6 skipped, the same skips as main), both at the head and at 8f6dda1 on its own. ruff check . (0.16.5, matching CI) is clean, and ops/leak-check.sh passes against main.
• Each new regression test was run against the pre-fix code and fails there.
• The verify job's own run: scripts, executed locally from the YAML:
  • pass against a wheel and sdist of this branch
  • fail against the published 0.1.1 wheel, on its version and separately on its loose schema
• The main-ancestry step passes the v0.1.1 commit and blocks an unmerged one.
• Built a wheel, an sdist and a wheel-from-sdist: each carries contextmesh/GRAPH.md and nothing at the top level. Imports cleanly from a venv outside the repo.
• CI on a141d08: 42/43 green, with the push-event copy of no private data skipped as designed (it runs on PRs only; the PR copy passed). That includes the release-candidate gate installing the built wheel on Linux and Windows, with and without [mcp].

Not verified

• Whether the pypi environment requires a reviewer. That setting isn't in the repo.
• The Windows leg of the publish workflow's verify job has not run anywhere yet. That workflow triggers only on release, so it first executes on the next release; the release-candidate gate's Windows wheel jobs, which cover the same install path, are green.

This branch was reset from main. It previously held #15's commits, which were superseded by #16 and closed; they remain under #15's refs.

🤖 Generated with Claude Code