feat(review): add fail-closed local review cascade (clean replacement for #26)

[aurascoper]

Clean replacement for #26, which had accumulated unrelated EEG research, Android
operations, and infrastructure work and was no longer reviewable as one change.

What this contains

The minimal tested dependency closure for the fail-closed local review cascade —
the first four commits of the original branch, with their identities preserved:

c007c2b  feat(logging): add local dialectic review
6ae9b96  fix(logging): fail closed on review citations
76a90e2  feat(review): add fail-closed local review cascade
2a8869e  docs(review): record local review governance

No cherry-picking or path surgery was needed. On the original branch the seven
unrelated commits were appended after these four rather than interleaved with
them, so the review cascade is a contiguous prefix and this branch is a lossless
extraction of it.

Diff scope (12 files):

Scripts/local_open_weight_review.py
Scripts/local_review_noninterference.py
Scripts/quarantine_dialectic_corpus.py
Scripts/research_decision_register.py
Scripts/review_quarantined_dialectics.py
Tests/eval/test_local_dialectic_review.py
Tests/eval/test_local_open_weight_review.py
configs/local-open-weight-review-v0.json
docs/architecture/dialectic-corpus-quarantine.md
docs/architecture/local-open-weight-review-cascade-gate-report.md
docs/architecture/local-open-weight-review-cascade.md
docs/scoping/research-decision-register.schema.json

Scripts/quarantine_dialectic_corpus.py (landed in #25) is modified by exactly
one line — adding eligible_for_science: False to the DISPOSITION contract,
which the cascade depends on.

Deliberately excluded

Preserved on the original branch, not transferred here:

NeuralComposeEEG/... jepa_synthetic + JEPA/SVD experiment scoping
function-space foundations research package
docs/NeuralCompose-android-client-prompt.md
docs/termux/*  (canonical copies belong in neuralcompose-client)
docs/operations/three-node-workflow.md

Verification

Run against this branch:

py_compile (5 scripts)                        OK
JSON validity (2 config/schema files)         OK
git diff --check                              clean
python3 -m unittest Tests.eval.{test_local_dialectic_review,
  test_local_open_weight_review, test_dialectic_corpus_quarantine}
                                              25 passed / 1 error

Final verification

The structured-state prerequisite landed in #35. The branch was then merged
forward onto the current docs/eeg-methods-scope head without rewriting its
four review-cascade commits.

Focused review-cascade suite:

• 26 passed
• 0 failed
• 0 errors

Full Python CI:

• NeuralComposeEEG contracts: 53 passed
• Tests/eval: 86 collected, 3 explicitly deselected under issue Three pre-existing Tests/eval failures deselected by the new Python CI job #38,
  83 passed

The eval collection is 86 rather than staging's 62 because this PR adds the
review-cascade suites: 62 on the base, plus the 24 new tests in
test_local_dialectic_review.py and test_local_open_weight_review.py
(the focused set is 26, two of which already existed in
test_dialectic_corpus_quarantine.py).

Additional gates:

• py_compile: pass
• JSON validation: pass
• git diff --check: clean
• Swift build-and-test: pass

The three deselections are pre-existing statistical-contract questions tracked
in #38. No entire module or file is excluded because of them.

One test change was required by the new base

5117a9e repoints the shadow replay test's output under root/shadow/.

It had been writing its state manifest to root/manifest.json — the same path
test_pipeline._manifest writes the source manifest to — so every run
silently clobbered the fixture. #35's immutable-publication guard turns that
into an error rather than a silent overwrite, which is how it was found. The
identical collision existed in #35's own suite and was fixed there the same way.

[aurascoper]

Prerequisite chain for this PR is now open:

• feat(eeg): add fail-closed structured-state shadow artifacts #35 — commits neuralcompose_eeg.structured_state, the module test_synthetic_structured_state_replay_remains_shadow_only imports. Ready, 13 contract tests.
• fix(eeg): resolve NumPy trapezoid compatibility lazily #36 — NumPy trapezoid portability fix. Needed before Python CI can run on a modern NumPy.
• ci: run the Python contract suites #37 — adds the python-contracts CI job, satisfying the github_ci_runs_tests_eval: true gate for this PR going ready. Stacked on fix(eeg): resolve NumPy trapezoid compatibility lazily #36.

Once #35 lands, this branch needs its base advanced so the module is present, after which the full 26-test set should pass. Holding as draft until then.

[aurascoper]

Correcting one figure in the description before merge.

CI collected 86 eval tests, not the 62 I originally recorded:

Ran 53 tests in 8.166s ... OK                      NeuralComposeEEG contracts
collected 86 items / 3 deselected / 83 selected
83 passed, 3 deselected in 3.12s                   Tests/eval

62 was the staging baseline. This PR adds test_local_dialectic_review.py and test_local_open_weight_review.py, contributing 24 tests (the focused set is 26, two of which already existed in test_dialectic_corpus_quarantine.py). 62 + 24 = 86. The EEG count of 53 matches the local run exactly.

Description updated. Nothing failed — the number moved because the suite grew with this PR's own tests.