Testing plan: record-feedback writes markdown that fails the project's own markdownlint

Derived from bugfix.md and design.md, before tasks.md — each task's _Test:_ names a row of the matrix below. Authored at test-planning and completed at verification. See reference/testing.md.

This file is executable content. It names commands an agent will run, so review it like code. Credentials appear by reference only; this work item needs none.

Test matrix

#	Type	Applies?	Scope / what it proves	Where it runs
T1	Unit	yes	record_feedback emits no line that is emphasis and nothing else (R1.1), and an empty body produces no blank-line pair (R1.2)	uv run --project cli python -m pytest -q cli/tests/test_graph_integration.py
T2	Integration (scenario)	yes	at the design-approval gate, a recorded approval keeps the handle and the body (R2.1, R2.2) and lands lint-clean by shape (R1.3)	uv run --project cli python -m pytest -q cli/tests/test_graph_integration.py
T3	Contract (OpenAPI / GraphQL SDL)	n/a — no API surface changes; the control-plane spec under docs/api-specs/ is untouched
T4	End-to-end	n/a — an end-to-end run needs a live webhook, a tmux session and a human approver; the gate's exit chain is already covered at T2, which is the seam the defect lives on
T5	UI / visual	n/a — no user-facing surface (design.md §UI/UX)
T6	Snapshot	n/a — the artifact's exact bytes are asserted directly at T1/T2; a snapshot would restate them less legibly
T7	Performance / load	n/a — two string branches on a path that runs once per human approval
T8	Security / abuse case	yes	the authorization boundary this hook sits behind still holds: an unauthorized author's approval is not read, and a self-authored comment is not feedback (existing negative tests, re-run because the file changed)	uv run --project cli python -m pytest -q cli/tests/test_graph_integration.py
T9	Accessibility	n/a — no rendered UI
T10	Migration / upgrade	n/a — no persisted state, no config key, no schema; artifacts already recorded into by the old code are explicitly out of scope
T11	Manual exploratory	n/a — replaced by T12, which is the same check made reproducible
T12	Linter conformance (added)	yes	the real linter, at the version make lint pins, accepts an artifact the fixed hook recorded into — and rejects the same artifact recorded into by the old hook (R1.3)	npx --yes markdownlint-cli2@0.18.1 "<artifact>"
T13	Repository gates (added)	yes	the whole repository still passes what CI runs: ruff, ruff format, pyright, config validation, the full suite, and markdownlint over every **/*.md — including the four spec artifacts this work item adds	make check

Scenarios & requirement trace

Row	Requirement(s)	Scenario / case
T1	R1.1	test_a_recorded_review_never_writes_emphasis_alone_on_a_line
T1	R1.2	test_a_comment_with_no_body_is_recorded_without_a_blank_line_pair
T2	R1.3, R2.1, R2.2	Scenario: an approval carrying suggestions is recorded in the artifact (extended: the recorded block is lint-clean and the body is verbatim)
T8	bugfix.md §Security considerations	Scenario: an unauthorized author's approval is not feedback · Scenario: the harness's own comment is not feedback (existing)
T12	R1.3	the fixed hook's output through markdownlint-cli2, and the old hook's output through the same, in the same run
T13	R3.1, R3.2	make check

Verification environment

• Repositories: this repo only.
• Services / containers: none.
• Fixtures & data: none beyond pytest tmp_path fixtures already in the suite.
• Credentials: none. This work item reads and writes no secret.
• Node.js: required for T12 and for the markdown half of T13 only — npx fetching markdownlint-cli2@0.18.1, the same invocation make lint uses. The Python suite (T1, T2, T8) deliberately does not depend on it.
• Bring-up: uv sync · Tear-down: none.
• If bring-up fails: record it under Verification results, leave the dependent activities unticked, and escalate — do not pass the gate on an environment that never came up.

Evidence plan

Row	Evidence	Path under evidence/
T1, T2, T8	the red run (both new tests failing against the unfixed hook)	red.md
T1, T2, T8	the green run, with counts	green.md
T12	both linter runs — old shape rejected, new shape accepted — with the exact output	shapes.md
T13	make check output, per target	check.md

Nothing captured here contains a token, a hostname or personal data: the fixtures use the literal author owner and bodies written for the tests.

Verification activities

• [x] T1 — uv run --project cli python -m pytest -q cli/tests/test_graph_integration.py
• [x] T2 — same command (the gate scenario)
• [x] T8 — same command, plus cli/tests/test_pdlc_e2e_integration.py in the whole-suite run
• [x] T12 — npx --yes markdownlint-cli2@0.18.1 over an artifact recorded by each shape
• [x] T13 — make check

Verification results

Activity	Command / procedure	Outcome	Evidence
T1	uv run --project cli python -m pytest -q cli/tests/test_graph_integration.py	pass — both new tests red before the fix, green after; 21 passed	evidence/red.md, evidence/green.md
T2	same command	pass — the extended gate scenario asserts the handle, the body verbatim, and no emphasis-only line	evidence/red.md, evidence/green.md
T8	same command, and the full suite for test_self_authored_and_unauthorized_comments_never_release_a_gate	pass — both negative tests unchanged and still passing	evidence/green.md
T12	npx --yes markdownlint-cli2@0.18.1 over old.md/new.md (the same two comments through each shape) and over the three candidate shapes	pass — 2× MD036 on the pre-fix shape, 0 errors on the fixed one; the ticket's blockquote candidate also passes, for the reason the design rejects it	evidence/shapes.md
T13	make check	pass — ruff, ruff format, pyright (0 errors), config validation, markdownlint over every **/*.md (0 errors), 2225 passed / 1 skipped	evidence/check.md

Not executed: none. Every planned activity ran.

Two findings from the verification itself, both recorded rather than smoothed over:

• This work item's own artifacts failed markdownlint on the first full run — an MD038 in bugfix.md (a code span written as `+ `) and an MD010 in shapes.md, where the captured linter output quotes a hard tab back. Fixed and disabled-with-a-reason respectively, before T13 was ticked. Worth stating: the gate that caught them is the same one this ticket is about.
• T12 confirms the ticket's blockquote suggestion works. It is rejected on the grounds in design.md §Chosen shape — it passes only because MD036 does not descend into blockquotes — not because it fails.

Review comments

Appended by the-loop's record-feedback hook when a human gate approves with comments (issue-109). Append-only and attributed: an approval never silently discards a reviewer's suggestions, and the feedback travels with the document it concerns rather than living in a side-channel tracker.