Status: COMPLETE — published and verified on 2026-09-06. The additive release preserves the public root exports and 21-tool MCP surface under the release policy.
| Milestone | Evidence | Disposition |
|---|---|---|
| Task-aware context and isolated demo | Foundation result | PASS |
| CLI with concurrent WAL clients | Concurrency result | 210 PASS, zero repairs |
| Portable evidence | Independent result | 71 PASS, including 100,003 records |
| Evaluated procedures | Independent result | 161 PASS |
| Bulk transaction boundaries | Independent result | 8 PASS |
| Scoped sharing | Original result, repair result | 193 PASS after an infrastructure retry; unchanged 193 PASS after composition repair |
The support change record separates historical frozen configuration from schema 7 and version 1.10.0. Original freezes, failing evidence and the explicit portable harness amendment remain unchanged. The 14 independently authored postimplementation composition probes supplement the frozen suite; they are not a new preimplementation freeze.
PR #89 merged reviewed head
ddffede03d87c4fb3c67550d25223ec7a430cff0 as
be9fdc408a040674d012dad60e2d0c4bd988ce05; their trees are identical.
The immutable annotated tag v1.10.0 points to that merged commit.
Final native CI
passed all nine jobs across Linux, macOS and Windows on Python 3.11–3.13,
plus Build Package and LongMemEval R@5 Gate. Each POSIX job passed 984 tests,
skipped 2 and deselected 8, with 83.31% coverage. Each Windows job passed 985,
skipped 1 and deselected 8, with 83.35% coverage. POSIX skips the Windows encoding
regression; every matrix job skips the release-tag test because its checkout
contains no release tags. The matrix selects not slow; the unchanged
100,003-record case passed separately in independent portable acceptance.
The separate model gate passed all 7 evaluations in 11.86 seconds.
Sanitized native evidence
records each job and its actual skips, counts and coverage.
The first CodeQL result flagged skipped base initialization in the scoped
facade. Repair 1 uses composition, explicit context cleanup and two adapter
typing casts. It opens no second database and changes no frozen assertions.
Final CodeQL
and its separate result check 101424916493 passed on the final reviewed head;
no alert was suppressed or dismissed.
All five jobs in Release workflow 34010797495 succeeded: verification, build, Trusted Publishing to PyPI, PyPI verification and GitHub Release. Version 1.10.0 is available on GitHub and PyPI.
The tagged release test gate passed 992 tests, including slow and evaluation cases, with 84.44% coverage. It skipped the Windows encoding case on Linux and explicitly deselected the self-referential publication-tag test; the subsequent PyPI verification job passed that release-tag gate after publication.
Independent publication verification matched both distributions byte for byte across the workflow artifact, GitHub Release and PyPI. Cryptographic verification pinned the certificate, release workflow, tag, source commit and release run attempt 1.
| Distribution | SHA-256 |
|---|---|
cairntir-1.10.0-py3-none-any.whl |
321a6275c52e09d3c7af44c0d4be878b4e176e0f3247621fb7bf653aa1f7b7e5 |
cairntir-1.10.0.tar.gz |
cc915da3afff5405cd9d4be23516327ed565a194bc6094223017663e6d41f2d5 |
Fresh public-package smoke
passed in an isolated Windows Python 3.11.9 environment. All 72 installed package
files matched the public wheel. Version, redirected help, recipes, v2
export/import/retry, abstention, recovery incompatibility and the demo passed.
Actual stdio MCP exposed 21 tools and agreed with CLI handoff; its complete
response used 1,851 characters under an 8,192-character ceiling. Doctor found
4 drawers and 4 vectors at dimension 512, SQLite integrity ok and no foreign-key
errors. The initial sandbox denied read-only host-configuration inspection;
the same package and harness passed with those reads allowed, without a runtime
change. The synthetic demo measured 7,720 versus 2,899 characters, a 62.45%
payload reduction, with no billing-savings claim.
The independent migration rehearsal preserved all 1,280 drawers and vectors, every original row across 19 tables, and the original table schemas. Schema 7 adds portable identities and procedure registries; the automatic backup matches the input snapshot. No embedding calls or source migration occurred.
The isolated candidate wheel reported 1.10.0 and exposed 21 MCP tools. CLI and
actual stdio MCP handoff agreed; the complete MCP response used 1,851 characters
under an 8,192-character ceiling. Abstention, recovery incompatibility, v2
export/import/retry, store integrity, redirected Windows help, recipes and the
synthetic demo passed. A legacy fixture was first migrated by normal status;
read-only task mode correctly refused to migrate it itself. These are candidate
checks; fresh published-package results are recorded separately above.
Local gate evidence retains the full regression run, report-encoding correction and affected-file rerun, 83.37% coverage, seven offline model evaluations and quality checks. Source review records the candidate archive and credential-pattern inventories.
Production launchers still resolve to installed 1.9.0. Publication does not replace that installation or reindex the live store. The migration rehearsal used a private online backup; package checks use fresh isolated environments. No new dependency, production configuration or license change occurred.
The installed production Claude launcher connected but its health command reported tools-fetch failure. A separately configured candidate launcher remains pending Claude’s project approval. Neither is counted as a passing candidate host test; actual isolated stdio MCP tests provide automated transport evidence. Project Claude/Codex/Cursor configuration and policy checks are ready. Production host restarts remain outside scope until an installation upgrade is authorized; Qwen is unavailable on this workstation’s command path.
Hosted preflight has no production store or configured vault, so it makes no production integrity or vault-comparison claim. Plugin metadata remains in the versioned repository; wheel/sdist inventories do not claim to contain it. Synthetic payload reduction is a measured local comparison, not a commercial host, billed-token or general task-success result.