Skip to content

Harden GNN admission, package boundaries and artifact validation - #33

Merged
docxology merged 5 commits into
mainfrom
work/aii-blockference-integration-20261115
Nov 17, 2026
Merged

docxology merged 5 commits into
mainfrom
work/aii-blockference-integration-20261115

Conversation

@docxology

Copy link
Copy Markdown
Collaborator

Scope

Preserve the original reviewed improvement lane, including GNN source/runtime admission, model-artifact validation, package boundaries and documentation. This is a draft review branch, not release or scientific approval.

Exact-head observed proof

At 7da818d14f367ebe4ee63c7ef97d2a2b3ca8b3ac, 17 installed-origin and refusal controls passed with no failures or skips. The strict validator used installed GNN 5.1.0 at 75c834bd8519dc7b5adc4d24f16bf393cfd1ecdf; parser and semantic checks passed.
GitVault test receipt 67632-aa3720ca66ef40559a41afcdca586033, SHA-256 20ddea1eb3eb943c9c3ebc9c83ff57146dc0cde1823839a628c7e1c2f5387cc0: source unchanged, reaped, no remaining descendants.

Earlier frozen-source quality, native wheel/locked-constraint build, isolated installation and outside-checkout import proofs are retained. They are not relabelled as new-head all-arm results.

Required holds

  • B06: research-maintainer scientific-definition/prose decision remains open.
  • B08: independently reviewed hosted GNN bootstrap and complete source-manifest delivery remain open.
  • Remote tag resolution was not requested in this local installed-origin proof; its result remains unresolvable-offline.
  • The repository-specific all release gate, independent review and required hosted checks remain prerequisites before promotion. No generic GNN status, release acceptance, scientific result or merge approval is claimed.

- Provision and verify authentic immutable bootstrap and version-correct full manifest before the original hosted all gate.
- Preserve original captured command failure output and exact external-document diagnostics controls.
- Reconcile only the observed current-main locked dependency delta; retain producer, science and full-release boundaries.
- Regenerate only the original full/fast/slow collection rows from the named producer.
- Preserve all historical timings, coverage, baseline negatives and measured/scientific source identities; no measurement imports or rebaseline.
@docxology

Copy link
Copy Markdown
Collaborator Author

Local CI reproduction, with current main 869d84a merged locally (merge commit f7f7ac3, not pushed); uv 0.13.7, GNN runtime from specs/gnn-pin.json (75c834bd), reviewed bootstrap and manifest resources (SHA-256 verified):

  • Lint and mypy: pass.
  • release_check.py --arm test-full on 3.11 (the required test (py3.11) job): 260 passed, 49 skipped, but coverage 88.53% < 90% floor. FAIL.
  • --arm all on 3.12 with GNN_RUNTIME_PYTHON, GNN_BOOTSTRAP and GNN_SOURCE_MANIFEST set: 264 passed, 45 skipped, coverage 88.50%. FAIL.
  • Baseline: origin/main 3.11 test-full passes at 91.00% (114 passed, 1 skipped).

The regression comes from the new blockference/gnn_source.py (137 statements, 22% covered). Its admission tests run in worker subprocesses, so coverage doesn't measure them, and the corruption cohort skips unless GNN_TEST_SOURCE_ROOT is set (CI doesn't set it). Setting it doesn't raise the measured coverage. Left as draft: the fix needs in-process coverage of gnn_source (or subprocess coverage), not a lower floor.

🤖 Generated with Claude Code

docxology and others added 2 commits November 17, 2026 11:15
Add 26 in-process tests for blockference.gnn_source: a hash-bound
test-double bootstrap and a minimal double gnn package exercise the
worker command, resource lookup, bootstrap hash/pin/session guards,
source-path restoration, probe and importer checks, and the main exit
codes (0/1/2) including refusal that keeps previous output. The module
reaches 99% line coverage and the suite 91.97% on CPython 3.12
(92.01% on 3.11). Regenerate the gate-metrics collection counts.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@docxology

Copy link
Copy Markdown
Collaborator Author

Restored the 90% coverage floor with in-process tests and qualified the full gate locally (hosted Actions unavailable). Head 4b5e0f9:

  • ef7d6f7: merged origin/main (869d84a, chore(deps): bump tornado from 6.6 to 6.6.1 #32), clean.
  • 4b5e0f9: added tests/test_gnn_source_worker.py, 26 in-process tests for blockference.gnn_source.
    • They use a hash-bound test-double bootstrap (BOOTSTRAP_SHA256 is pinned to the double's digest, so the real hash check still runs) and a minimal double gnn package.
    • Covered paths: worker command, resource lookup, bootstrap hash mismatch / changed bootstrap / single-session guard, missing or mismatched pin, pin report, JSON coercion, _source_path restoration, probe in source mode, installed-mode import failure, importer check folding, and main exit codes 0/1/2, including a refusal that keeps previous output.
    • gnn_source.py is at 99% line coverage.
    • Also regenerated docs/_generated/gate-metrics.md collection counts (335 full / 318 fast / 17 slow).

Local gate (same steps as the CI jobs):

  • release_check.py --arm test-full on CPython 3.11: 286 passed, 49 skipped, 92.01% coverage (90% floor).
  • release_check.py --arm all on CPython 3.12, with the GNN 5.1.0 runtime at pin 75c834bd plus custody bootstrap and source manifest: lint, docs/gate-metrics check, test-full (290 passed, 45 skipped, 91.97%), manuscript, gnn validation (tag v5.1.0 and revision verified), smoke and package all pass. Result: "release gate passed".
  • ruff check tests/ is clean.

🤖 Generated with Claude Code

@docxology
docxology marked this pull request as ready for review November 17, 2026 20:22
@docxology
docxology merged commit d6599fa into main Nov 17, 2026
0 of 4 checks passed
@docxology
docxology deleted the work/aii-blockference-integration-20261115 branch November 17, 2026 20:22
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