Verification — issue-321

The testing plan executed (testing-plan.md, rows T1, T2, T8, T10, T12). Commands run from the repository root at the head of claude/github-issue-321-tipucc. Every token in the tests is a fixture (xoxb-test); the checkouts are git init temp directories with a https://github.com/o/r.git origin. Nothing here needed redaction.

The reproduction, before the change

A scratch script against ba0c433 (13.3.0): a real checkout with origin, a registry record whose cwd is that checkout, graph-state.json parked at requirements-approval, a start recorded in the portable control store, the default control policy. The dispatcher's coupling and the pipeline's read, over the same config:

dispatcher sees gate: True
pipeline sees gate: False

Red → green, per task

The two T2 scenarios (task 1) were written first and run against ba0c433:

uv run --project cli python -m pytest -q cli/tests/test_bus_integration.py -k "default_control_policy or no_session_record"
E       AssertionError: assert (1 == 1 and [('github:o/r#7', 'approved')] == []
E         Left contains one more item: ('github:o/r#7', 'approved')
FAILED cli/tests/test_bus_integration.py::test_an_approval_from_slack_reaches_the_gate_under_the_default_control_policy
FAILED cli/tests/test_bus_integration.py::test_a_reply_for_a_work_item_with_no_session_record_is_left_to_the_ledger
2 failed, 8 deselected in 0.17s

The failure is the bug in one line: the pipeline classified the approval as a work-item.reply and delivered it into the session (the marked mirror, invisible to the gate) instead of leaving it, unmarked, to the ledger.

Task	Red (before the change)	Green
1 reproduction	2 failed (above)	10 passed (test_bus_integration.py)
2 reader	_graph_reader absent; _at_human_gate returned False for a missing record	test_the_pipeline_reads_the_graph_through_the_dispatchers_own_coupling, test_no_session_record_is_an_unknown_gate_not_a_closed_one, test_a_disabled_graph_coupling_means_no_gate_to_answer, test_the_pipelines_read_moves_nothing, test_a_graph_fault_is_cannot_tell
3 classification + record + event	classify took no grants; channel.reply_received carried no gate	the remaining six T1 tests
4 docs	—	make lint (markdownlint over 965 files), test_docs_parity.py

Rows T1, T2, T8, T10

== T1
73 passed in 0.45s
== T2
10 passed in 0.20s
== T8
5 passed, 68 deselected in 0.07s
== T10
8 passed in 0.12s

Row T12 — make check

make lint:

uv run ruff check cli hooks
All checks passed!
npx --yes markdownlint-cli2@0.18.1 "**/*.md"
markdownlint-cli2 v0.18.1 (markdownlint v0.38.0)
Linting: 965 file(s)
Summary: 0 error(s)

make format-check, make typecheck, make validate:

uv run ruff format --check cli hooks
277 files already formatted
uv run pyright cli
0 errors, 0 warnings, 0 informations
uv run python scripts/validate_config.py
VALID   .the-loop/harness-config.yaml
VALID   skills/the-loop/templates/harness-config.yaml
VALID   cli/the_loop/harness-config.default.yaml
VALID   .the-loop/collaborators.yaml
VALID   skills/the-loop/templates/collaborators.yaml
VALID   .the-loop/cli-config.yaml
VALID   skills/the-loop/templates/cli-config.yaml

make test (the full suite):

uv run --project cli python -m pytest -q cli
3033 passed, 1 skipped in 156.64s (0:02:36)

The first make lint run found two unused imports in the new tests (ruff F401) and the first format-check two files to reformat; both fixed and re-run, as recorded above. No existing assertion was changed: the eight issue-309 scenarios that patch _at_human_gate to True / False pass unchanged against the three-valued read.