Bump mlx from 0.29.3 to 0.29.4 in /tools/mlx-smoke - #97

mlx-lm 0.30.2 pinned transformers==5.0.0rc1, which Dependabot alerts 3-6 flag
(GHSA-xrqw-3rrv-vx5w, GHSA-fgcw-684q-jj6r, GHSA-29pf-2h5f-8g72,
GHSA-69w3-r845-3855), and the lock resolved mlx 0.29.3 (GHSA-j842-xgm4-wf88,
GHSA-w6vg-jg77-2qg6). requirements.in now pins mlx-lm 0.31.3 with floors
transformers>=5.10.0 and mlx>=0.29.4; the lock (same cutoff) resolves mlx
0.32.2, mlx-lm 0.31.3 and transformers 5.17.0. Newer mlx wheels are macOS 14+,
so the lock and the CI diff-check both compile with MACOSX_DEPLOYMENT_TARGET=14.0.
Supersedes Dependabot #97 and #98, whose raw lock edits fail the diff-check.

Co-authored-by: seathatflowsinourveins <[email protected]>
Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>

…ison's confirmed defects and file its evidence (#121)

* Fix the closed automation catalog's confirmed defects and file the blind-comparison evidence

Correct the stale harden-runner ubuntu-job count (18 of 22 -> 20 of 24,
measured at HEAD with the exact counting method), fix dependabot.yml's
"github-actions only" header and every doc claim that Dependabot owns only
GitHub Actions references now that a deliberately vulnerable pip fixture
entry exists, and fix the decision record: the stale "only the osv-scanner
job holds security-events: write" claim (zizmor-sarif-upload also holds it),
the "separate upload job scans twice" rationale that conflicts with the
chosen design, and add the missing Evidence/Alternatives/Decision/Overturn
elements to the zizmor-online, Scorecard SARIF and dependency-review
sections. Add mutation-checked assertions covering the OSV/zizmor SARIF
upload failure-path semantics to tests/test_workflow_hardening.py.

Record the "Blind comparison (2026-09-23)" section in the decision record
and copy its comparison artifacts into
evidence/artifacts/blind-comparison-20260923/ (with the judgments file's
scratch/home paths redacted and a README listing every file's sha256),
registered in manifests/evidence.json.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

* Resolve reviewer findings on the GitHub-automation closure polish

Fix the review's confirmed high/medium findings:
- security-scan.yml section 3 never had only one job holding
  security-events: write (zizmor-online held it from the founding commit
  a9e5180); say so, with the verification command that shows the count.
- Restate the OSV separate-upload-job rejection honestly: it would not
  scan twice, and the real trade-off is the write scope sitting on the
  job that runs a curl-installed binary; mark it keep-but-compare.
- Reword the dependabot.yml header and the two docs/github-automation.md
  passages: ignore: urllib3 suppresses security-update PRs for the
  fixture pin; open-pull-requests-limit: 0 only stops version updates;
  alerts still appear via the dependency graph regardless.
- Restore the sealed 2026-09-22 lane:claude Dependabot why_not_default
  text verbatim in catalogs/landscape/foundation.json, and move the
  pip-fixture correction into a labelled coordinator annotation in the
  layer's limitations list instead of overwriting lane output;
  regenerate layer-verdicts-20260922.json and the handbook with
  build_verdicts.py --write.
- Section 5: cite the observed hosted upload (the first push run on
  4970ba0, dispatch run 35824151483) and scorecard-action's documented
  example workflow at its pinned SHA, not just the source diff.
- Section 4: name zizmor's actual online-only audit classes instead of
  claiming offline analysis covers no secrets class (its secrets audits
  are static). Section 6: replace the #97/#98 "nothing to flag"
  inference with what those PRs actually show (gate latency, not
  threshold behavior).

Fix the low findings:
- tests/test_workflow_hardening.py: the failure-path test read a fixed
  line offset after each step/job marker, so a reordered key or a
  folded if: block scalar could pass vacuously. Add step_block()/
  block_if() helpers that search the whole block and resolve folded
  scalars, and assert zizmor-online's analyzer step carries no
  continue-on-error (a crash there could otherwise still upload a
  partial SARIF). Mutation-verified: reordered uses:/if: with the
  cancelled() guard dropped, added continue-on-error to the analyzer
  step, and folded the upload job's if: with a hidden always() --
  each mutation failed the test before being reverted.

Re-registered all seven changed files in manifests/evidence.json with
scripts/host_receipts.register_file.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

* Split the OSV SARIF upload into a tool-free job; resolve the polish review findings

The osv-scanner job held security-events: write while running the installed
scanner on every event; it is now contents: read and a needs/!cancelled()
osv-sarif-upload job (no shell step or installed tool) uploads, matching the
zizmor split. Record sections 3 and 5 are rewritten to match (Scorecard upload
now cited from the code-scanning analyses list). The lane-sourced catalog text is
restored verbatim and its staleness recorded instead; the redacted judgments copy
now replaces only path tokens; the folded-if bypass in the artifact-step test is
closed. Mutation-checked: a folded always(), a lost !cancelled() and a write scope
back on the scan job each fail the suite.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

* Resolve the final polish review: re-run-safe artifact names, bypass-proof step checks, corrected citations

Artifact names now include run_attempt so a re-run cannot hit a same-name
conflict. step_block() spans the whole list item and block_if() collects
plain-scalar continuations, so a leading '- if:' or an always() on a continuation
line is caught (mutation-checked). Docs: harden-runner 21 of 25 ubuntu jobs after
the split; Scorecard layout cited to the example the README links
(ossf/scorecard scorecard-analysis.yml @ d13ba3f) and the upload to paginated
analysis ids; the unmerged-branch commit dropped from the grep list; the
redaction and upload-job wording corrected.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>

---------

Co-authored-by: seathatflowsinourveins <[email protected]>
Co-authored-by: Claude Opus 5.5 (1M context) <[email protected]>