fix: stop an arriving code answer from freezing the pane for its length

ROADMAP.md

@@ -348,7 +348,7 @@ Every item on the deferral shelf below has a milestone here, so a milestone-filt
 | ✅ | `SHEET_KEEP` names two things, so removing a gesture reassigns the never-drop set by rank | [#464](https://github.com/breferrari/vigia/issues/464) |
 | ✅ | A release that removes a key can be filed as internal, and 0.42.0 was | [#467](https://github.com/breferrari/vigia/issues/467) |
 | ✅ | A hidden path still wakes the watch. **Deferred by [#495](https://github.com/breferrari/vigia/issues/495): it changes what I1 measures** | [#501](https://github.com/breferrari/vigia/issues/501) |
-| ⬜ | A note's rows are built whole before the clamp trims them. **Deferred by [#452](https://github.com/breferrari/vigia/issues/452): the steady frame is 12.557ms against 12.483ms as prose, and it is the frame an answer arrives on that scales** | [#530](https://github.com/breferrari/vigia/issues/530) |
+| ✅ | A note's rows are built whole before the clamp trims them. **Deferred by [#452](https://github.com/breferrari/vigia/issues/452): the steady frame is 12.557ms against 12.483ms as prose, and it is the frame an answer arrives on that scales** | [#530](https://github.com/breferrari/vigia/issues/530) |
 | ✅ | The caret row's weight is the one modifier a theme file cannot reach | [#195](https://github.com/breferrari/vigia/issues/195) |
 | ⬜ | Three overlays are kept apart by hand in eight places. **Deferred by [#511](https://github.com/breferrari/vigia/issues/511) and again by [#508](https://github.com/breferrari/vigia/issues/508): a refactor of the two that predate the branches that found it.** [#519](https://github.com/breferrari/vigia/issues/519) is its duplicate, filed before this row was found and closed against it | [#515](https://github.com/breferrari/vigia/issues/515) |
 | ⬜ | The sheet's tables are audited, not derived, so the keymap can still drift into them | [#312](https://github.com/breferrari/vigia/issues/312) |
@@ -580,6 +580,7 @@ Items that moved into an *earlier* phase than planned. Recorded for the same rea
 | Exclude files frozen at watch start ([#621](https://github.com/breferrari/vigia/issues/621)) | Off the Shelf, 2026-09-30 | Filed by #619 and still real. Two stats a tick while events flow, none while idle. |
 | Mixed release notes fail the bump ([#624](https://github.com/breferrari/vigia/issues/624)) | Off the Shelf, 2026-09-30 | Taken by the unattended loop, whose merged #617 and #622 it breaks the next release on. Workflow work, which the loop takes when it blocks releasing the loop's own PRs. |
 | The heat projection's cost follows the file ([#55](https://github.com/breferrari/vigia/issues/55)) | Off the Shelf, 2026-09-30 | Taken by the unattended loop when `next.sh` ran dry. The row's reason was that nothing had ever taken it, which is not a reason that can hold. The fix is a slot beside the cached diff: a file's heat strip is projected once per computed diff and served from the frame while that diff is reused, so a frame pays a copy per drawn heading and the walk over the hunks is paid once per computed diff |
+| A note's rows are built whole before the clamp trims them ([#530](https://github.com/breferrari/vigia/issues/530)) | Off the Shelf, 2026-09-30 | Taken by the reader's approval of the plan when `next.sh` ran dry. The row's reason was a cost statement, not a blocker: the arrival frame of a fenced answer scaled with the answer's length and was paid once. A note is now counted whole and built only where the clamp keeps it, and a quoted block is parsed as far as the last line the frame draws, so an arrival costs a screenful |
 
 ---
 

RULINGS.md

@@ -185,6 +185,18 @@ The middle column **is** the floor. One real sibling pays the compile in full. T
 >
 > **Fixed 2026-09-04 ([#412](https://github.com/breferrari/vigia/issues/412)), which makes the line above false.** Every fresh diff now puts its height beside it, and the walk asks only the span, with one stat. A print that moved inside the settle margin keeps its height until the file settles, and the walk reads it once then. Re-reading inside the margin was #84's 20.71ms breach of I9. So the in-margin re-measure above became a deferral: one stat per file per tick, and one read when the file settles. That frame is asserted against I9.
 
+## I4 — a note's rows and a quoted block's parse follow the window, 2026-09-30
+
+Before [#530](https://github.com/breferrari/vigia/issues/530), `Pin::rows` built every row of every note twice a frame, once for the clamp's count and once to emit, and `answer_of` parsed every line of a fenced block on the frame the answer arrived. Measured in release on the note rig on 2026-09-12: a 2,000-line fenced answer cost 160.5ms on its arrival frame against 13.2ms as prose, and 10,000 lines cost 757.2ms. The steady frame held (12.557ms fenced against 12.483ms prose for twenty four-line answers), which is why the row sat on the Shelf as "paid once".
+
+The mechanism: a note is one list of sections (`Pin::sections`) that the count sums and the build walks, so the clamp's exact count and the emitted rows cannot differ, and a unit test (`view::tests::count_matches_build`) holds every shape, width and window to that. The window functions report the rows they walked, so a part is not counted before it is built. A quoted block keeps its parser position (`Quote::parse`) and `Pass::quoted_to` fills it forward to the last line the window reaches; `HighlightStats::quoted_lines` counts the lines parsed. `View::built` counts the note rows a screen built.
+
+Measured on the branch, release, with `arrival_holds_the_budget`: the 80x160 pane's fenced arrival went from **119.98ms p50 on `main` to 18.69ms**, against 13.14ms for the same bytes as prose, and it parsed 110 lines for 115 rows; the 80x24 pane reads **7.90ms fenced against 7.55ms prose**, 8 lines parsed for 13 rows. The tall pane's remainder is its window, a hundred-odd lines of grammar on top of a 13ms frame, so the gate times the ordinary pane and reads the tall one. The steady quoted gate moved from 12.557ms to 12.52ms fenced.
+
+Three alternatives were rejected. A lazy clamp that never needs an exact count: the bottom clamp computes `dropped = total - height` and a box pulls the window by row sums, so the count must be exact, and only the build can be windowed. Parsing at placement with a line bound from the layout: placement runs before the clamp decides the window, so the bound would be circular; the rows the clamp keeps are the only position that knows it. Sharing the hunk's `Entry::fill_to`: a hunk has checkpoints, a rewind and three parse states because a line of it changes before every frame; a quote has one `Side` and is replaced whole on rewrite, so the ten-line fill loop is the only shape shared and a shared type would carry the hunk's machinery for nothing.
+
+Left as it is: the bottom clamp on an answer taller than the pane parses to the last drawn row, which is the whole block, once, the same first touch the #51 bullet records for `G` into a hunk. The count walk is a width scan of the whole answer per frame, twice at most; `prose_pieces` rescans a paragraph's remainder per piece, which a 100 KB single line would feel, and which predates this change.
+
 ## I2 — why it is two numbers
 
 > [!NOTE]

SPEC.md

@@ -74,7 +74,7 @@ A regression past any budget **fails the build.**
 > **The measurement histories behind these rows are in [`RULINGS.md`](RULINGS.md)**
 >
 > - **I1** has a cost its row does not measure. `crossterm`'s mouse bundle sets `?1003h`, so pointer motion over the pane wakes the loop, and that cannot be reclaimed portably. **A batch that asks for nothing, moves no pointer mark and meets no deadline paints nothing**, so the wake is all it costs ([#154](https://github.com/breferrari/vigia/issues/154)).
-> - **I4** was **narrowed 2026-08-01**. A height is hunk boundaries and line counts, not content. That is the row's first exception, and it is why #49's refusal of a repository-wide total still holds for sums over content. The height walk became **incremental 2026-08-04** (#101). Since 2026-09-04, a file that moved inside the settle margin keeps its height until it settles. It is then read once, on a wake at the margin's end ([#412](https://github.com/breferrari/vigia/issues/412), [#425](https://github.com/breferrari/vigia/issues/425)). The numbers and three rejected alternatives are in the ledger.
+> - **I4** was **narrowed 2026-08-01**. A height is hunk boundaries and line counts, not content. That is the row's first exception, and it is why #49's refusal of a repository-wide total still holds for sums over content. The height walk became **incremental 2026-08-04** (#101). Since 2026-09-04, a file that moved inside the settle margin keeps its height until it settles. It is then read once, on a wake at the margin's end ([#412](https://github.com/breferrari/vigia/issues/412), [#425](https://github.com/breferrari/vigia/issues/425)). The numbers and three rejected alternatives are in the ledger. A note's rows and a quoted block's parse follow the window too ([#530](https://github.com/breferrari/vigia/issues/530)): a note is counted whole and built only where the clamp keeps it, and a block is parsed as far as the last line a frame draws, with the parser position kept beside the block. The frame a fenced answer arrives on costs a screenful of grammar, and a block drawn from its bottom costs the whole block once: on a tall pane either is a first touch outside I9, the shape the #51 bullet in §10 records for a hunk. Ruled 2026-09-30, reader. User-facing: an answer the agent quotes as code no longer freezes the pane for the length of the block when it arrives. Gated by `crates/vigia/tests/budgets.rs::arrival_holds_the_budget` on the ordinary pane and by `crates/vigia/tests/notes.rs::parse_follows_the_window` and `rows_follow_the_window` on the counters. The numbers and the alternatives are in the ledger.
 > - **I2** is two numbers because re-diffing (`gix`, Phase 1) and re-highlighting (`syntect`, Phase 2) have different dependencies and phases. The 18.58ms-vs-3.27ms measurement that forced the split is in the ledger.
 > - **I8** no longer names `SIGINT` for Ctrl-C, because raw mode makes Ctrl-C a key event, never a signal. The externally delivered `SIGINT` is **covered since 2026-08-08** ([#24](https://github.com/breferrari/vigia/issues/24)). It was taken for both tier-1 platforms in one issue, because the single-task version was Unix-only and [#16](https://github.com/breferrari/vigia/issues/16) had already rejected a guarantee that means different things on different targets. Measurement made that affordable: neither half adds a crate to any graph. The row's exclusions now read the same on both platforms.
 > - **I3**'s scheduled soak runs 30 minutes a week on Linux over the 20 x 200 fixture. The fixture settles in about thirty seconds, so the window measures the plateau. *Ruled 2026-09-29, reader.* The full 24h runs by `workflow_dispatch` or locally. The sample count is fixed, so the statistic is computed the same way at any window.