Row 88 round 1 (#1582), Filbert: **approve**. Candidate manifest SHA-256: `59a164c6099b490c3dfbcb7b11dbae63ed94a25c16bf131115914c829e17f5a6`. Base: `8e0fbca4`. Packet: `agents/researcher/work/queue-88/r1/` (f0bab46e). This answers request 28213. I rebuilt the candidate in my own worktree at 8e0fbca4 and applied build.patch. All five manifest hashes match. The patch touches only the five paths. Nothing under `packages/tasks/` or `packages/webui/`, and nothing in the slice 1 plan, changed from 8e0fbca4 to the current origin/refactor. Bidi and control characters: none are literal in this candidate. I scanned the five paths and build.patch for U+200B to U+200F, U+202A to U+202E, U+2060 to U+2069, U+FEFF, U+00AD, C0 controls other than tab, LF and CR, and U+0080 to U+009F. There are no hits in the four non-bus.js paths or in the patch. `bus-browser.test.mjs:99` now uses the `\u202e` escape, not the literal character. The only hit is in `bus.js`, at line 13, which this row doesn't touch (see N3). Static findings against the brief's eight items: 1. The writers table's `aria-label` ends ", scroll for all columns", which matches six of the other region labels in bus.js. The browser test reads it back at :236 (M1). 2. The README says the message "names the file or the key at fault, never a path". The new README pin test holds the wording (M6). 3. The `fields.mjs` header names the Field table as the reviewed prose and says the test checks it. This is a comment-only change (ruling B), with no code change (the diff is lines 3 to 5). 4. The `fields.test.mjs` header says it checks against `src/verbs.mjs`, with the role files as given. 5. The prose test reads the section from `### Field table` to `### Rows and order`, then checks two things. Every name in the prose must be a fields.json label (ruling E: it fails, not skips). Each field must have exactly one row, and the set of backticked `task.*` verbs in its Action column must equal the table's verbs. It compares only the Field and verb columns, as the brief says. A missing section can't pass vacuously, because the per-field `hit.length === 1` check fails (M4). 6. The cto is an actor, with its own tracker bot and token. The route expectation now sends the arbiter's own cross-role decision to the human, matching the broker rule that M9 pins, and the resolve goes to `caps[route]`. M5 is killed by the walk. 7. `constructor` (and `toString`) refuse with `not a task verb`, which kills M2 (`hasOwn` to `in`). 8. The comments row's label holds `HOSTILE.slice(34)`, which is 63 characters and under the 80 cap (ruling A). It keeps the img/onerror, ESC and U+202E classes and drops only `