fix(viewer): the span gutter moves to the right edge; any span's intent promotes to a comment

[folded]

The strip moves to the right edge

The centre gutter pushed the two code columns ~250px apart and broke the
side-by-side read. Line numbers stay on the left of each half — they are a
lookup target, and tracing a line to a number across whitespace is hard — so
the strip goes to the outside:

 12 │ old code        │ 12 │ new code        ▌ rationale

New half grid is lineno · code · bars · text, the strip sticky to the right
edge — the plain mirror of the line numbers' sticky-left (right: 0 on the
text, right: var(--gutter-text-w) on the bars). Bar against the code;
innermost bar nearest the code (depth 0 against the text, each level 6px
nearer). The .diff split is untouched, so the code viewports stay equal:
424.406 / 424.422 at 1440px, 254.406 / 254.422 at 1100px. Scrolling the half
80px moves the code −80 and the strip 0.

Collapsed (g) is bars alone, flush with the half's edge; both code viewports
widen equally; no row stretches; hover, click, reload and panel as before.

Promote any span's intent

#34 retired the code-side note form and with it "Add as comment" on a span's
intent — which had only ever existed for single-line spans. Every block's pill
row now carries + comment, so multi-line spans get it for the first time.
Reuses the smell-pill path (promote({derivedFrom, …})); a promoted intent
lands as a comment on the span's first line.

Promoting a multi-line span hides its text block and keeps its bar — the
comment stands in for the text, but the bar marks a range the comment does not.
The single-line rule (dot goes too) is unchanged. Both sites that must agree do:
_attachSpans draws a promoted span's bar but no block, and the comment sweep
removes .span-text and dots only. openPromotionEditor and the editor's
prefillBody/derivedFrom were dead since #34 and are removed.

Mutations verified

Un-stick the strip (moves −80 with the code) · bars from the code side (depth 0
lands nearest the code) · drop both takeRecords() (page never settles — the
MO→pass→MO loop starves render; with the guard: 13 RO callbacks per toggle for
13 halves, 0 MO) · affordance on single-line blocks only (3 fail) · sweep
removes bars (2 fail).

Two things worth a follow-up, not widened into here

Comment boxes under the strip. .cell-annotation spans every column, so in a
hunk whose lines overflow the viewport a 64ch comment box runs under the sticky
strip. Confined the new half's annotation cell to grid-column: 1 / 3 so boxes
wrap before the strip whenever the row fits; a full fix caps .annot-box at the
visible code width (container-type: inline-size on .half).

The pill row costs a line on smell-less blocks — a span with no smells now
starts its intent one line below the bar's first row, because the + comment
button lives in that row. Alternative: float the button inside the intent when
there are no pills. Followed the brief's wording; flagging for the reviewer.

Pre-existing, noted: Render.renderHunkReplace (SSE hunk) swaps the hunk
DOM without Comments.renderAll(), so that hunk's threads vanish until the next
full render. Not introduced here.

Gate

414 vitest (410 → 414), 1051 Python unchanged, pyright 0, ruff clean.