Testing plan: one verbosity switch over the whole trace, defaulting to quiet

Derived from bugfix.md and design.md, before tasks.md. Authored at test-planning, completed at verification.

What is testable here, and what is not

Everything this change adds is pure or rendered, and both are already driven by the suite: model.test.ts calls exported functions directly, and App.test.tsx drives the whole dashboard against the demo transport. So the matrix is automated apart from one row — how the quieter panel looks — which needs a stylesheet jsdom does not apply. That row is a browser pass on the dev server, recorded with a screenshot.

Test matrix

#	Type	Applies?	Scope / what it proves	Where it runs
T1	Unit	yes	`isReadable` over every `ThreadRow` kind (malformed kept, tool result dropped, meta by its text, a tools-only assistant turn dropped) and `isBookkeeping` over the classification table (each hidden family, each kept family, `level=error` in a hidden family, an unclassified family, an empty event name)	`bun run test src/api/model.test.ts`
T2	Integration (scenario)	yes	`TranscriptView` and `EventTrail` rendered both ways round: quiet drops the tool groups, the empty meta rows and the plumbing events; verbose restores every one; the error event survives quiet; the hidden-count line appears only when filtering emptied a non-empty source	`bun run test src/components/Transcript.test.tsx`
T3	Contract (OpenAPI / GraphQL)	n/a — no API surface changes; the panel reads the same two routes with the same fields.
T4	End-to-end	yes	The dashboard on demo data: the switch is off at load, the stream carries the agent's prose and the malformed line and no tool group, and turning it on brings the tool calls back — from the mouse and the keyboard	`bun run test src/App.test.tsx`
T5	UI / visual	yes — manual, browser	The quiet panel reads as a conversation rather than a log, and the hidden-count line sits where an empty state would	`bun run dev`, screenshot into `evidence/`
T6	Snapshot	n/a — the rows are asserted by their text and their `data-entry` / `data-tools` hooks in T2 and T4, which is the snapshot with the reasons attached.
T7	Performance / load	n/a — two predicates over the ≤200 rows the routes already cap; no measurable budget.
T8	Security / abuse case	yes	The abuse case from `design.md` § Security design: an event in a hidden family carrying `level=error` still renders, so nothing hides itself by its name	`bun run test src/api/model.test.ts src/components/Transcript.test.tsx -t "error"`
T9	Accessibility	yes	The switch keeps `role="switch"`, an `aria-checked` that tracks its state, and an `aria-label` that is no longer `Tool calls`; `Enter` and space both toggle it	`bun run test src/App.test.tsx`
T10	Migration / upgrade	yes	Nothing persisted changes: `the-loop:settings:v1` gains no key, and the whole existing UI suite still passes unchanged apart from the three tests that asserted the old copy and the old default	`bun run test`
T11	Manual exploratory	yes	Against a real service with a long-polled work item: the trail that motivated the issue, before and after	an operator's machine; not available in this environment
T12	Documentation parity	yes	The control-plane capability doc describes the switch by its new name and scope; markdownlint passes	`make lint`, or its markdownlint half directly where `uv` cannot run

Scenarios & requirement trace

Row	Requirement(s)	Scenario / case
T1	R1.2, R1.3	`Scenario: a stream row is readable when it carries prose, thinking, a summary or a drifted line`
T1	R1.4	`Scenario: an assistant turn whose only content was tool calls is not readable`
T1	R1.2, R2.2	`Scenario: an event is plumbing when its family is, and is kept when nobody classified it`
T1, T8	R2.1	`Scenario: an error-level event in a hidden family is never plumbing`
T2	R1.1, R1.2	`Scenario: the quiet stream drops the tool groups, the empty meta rows and the plumbing events`
T2	R3.1	`Scenario: the verbose stream renders every row the quiet one hid`
T2	R4.1, R4.2	`Scenario: a trail emptied by the filter names its hidden count and the switch`
T2	R4.3	`Scenario: a source that is empty for its own reasons keeps its own empty state`
T4	R1.1	`Scenario: the trace panel loads with the switch off and no tool group in the stream`
T4	R3.1, R3.3	`Scenario: turning the switch on brings the tool calls back, from the mouse and the keyboard`
T9	R3.2	`Scenario: the switch names the whole stream's verbosity, not the tool calls`

Verification environment

The ui/ workspace on this machine: bun install --frozen-lockfile, then bun run test (vitest + jsdom + Testing Library), bun run lint (oxlint, type-aware) and bun run build (which runs tsc --noEmit first). These are the same three commands .github/workflows/ci.yml's ui job runs, so local equals CI. The Python side is untouched; make check is run once to prove it.

Evidence to capture

Under docs/specs/issue-419/evidence/:

verification.md — each row of this matrix with its command, its outcome and the counts from the run.
self-review.md — the review cycles and every finding's disposition.
security-review.md — the abuse case above, and the fail-open argument.
documentation.md — the capability doc updated in this PR.
the T5 screenshot, or the reason it did not run.

Verification results

Recorded at the verification node in evidence/verification.md: T1, T2, T4, T5, T8, T9, T10 and T12 pass; T3, T6 and T7 are n/a as planned; T11 did not run for want of a live service, and the Python half of T10 (make check) could not run because this container's uv predates the version pyproject.toml requires — the change touches no Python, and CI runs it on the pull request.

Activities checklist

[x] T1 unit tests written red, then green
[x] T2 component tests written red, then green
[x] T4 / T9 dashboard tests updated for the new default and copy
[x] T10 full bun run test green
[x] T5 browser pass, or its absence recorded with a reason
[x] T12 markdownlint green over the changed Markdown and the capability doc updated (make lint itself could not run here — see the verification record)
[x] Evidence committed