fix(layout): evaluate (when FORM . SPEC) display specs and measure buffer :align-to from the text area

[tag-und-nacht]

Dashboard's centered banner (and any buffer text using (space :align-to REGION) or a (when FORM . SPEC) display spec) lands where GNU Emacs 31.0.90 puts it. Two defects, both visible on the dashboard logo:

1. :align-to region targets were measured from the window edge, not the text area. center/left/right in buffer text came out one left-fringe width too far left (8px on a default frame). GNU produce_stretch_glyph makes a resolved region coordinate text-area-relative for buffer text and only then subtracts the pen (src/xdisp.c:32853-32859); chrome rows keep window coordinates and are unchanged.
2. (when FORM . SPEC) was applied whenever FORM was non-nil. Dashboard's line-prefix carries (when (display-graphic-p) …) and (when (not (display-graphic-p)) …); a graphic frame took the text-terminal clause (42px off with a 17px font). GNU takes nil/t literally and otherwise evaluates FORM through dsafe_eval with object, position and buffer-position bound, an error counting as nil (src/xdisp.c:6130-6160); a list of specs stops at the first replacing element for strings and lets later ones override for buffer text (6034-6040, 6055-6061).

How the when forms are evaluated

The layout walk cannot run Lisp while it holds the buffer, so the forms of the span the walk can reach (the one fontification just covered) are evaluated once before it:

• Context::display_when_form_holds (neovm-core, display/display_when/) is the dsafe_eval port: literal nil/t, inhibit-redisplay/inhibit-quit/inhibit-debugger bound like safe_funcall, the three GNU bindings, signals muted to nil, other flows propagated. Context::evaluate_display_when_sites roots every site's form and object (precise collector) and runs with the window's buffer current.
• evaluate_window_display_when_forms (layout engine, display_when.rs) collects the sites: display/line-prefix/wrap-prefix text and overlay properties, the buffer-local and default prefix values, the display properties inside prefix strings, and overlay before/after-strings placed by the walk's own resolver (RustTextPropAccess::overlay_strings_at, so the window filter, the invisible-text rule and the load position have one owner). The scan is closed at its end so strings anchored at point-max are reached (GNU next_element_from_buffer, src/xdisp.c:9846-9860).
• Results ride in LayoutBufferSnapshot and reach every reader through LayoutBufferView::layout_display_when_conditions: the text cursor, the row route, the bridge predicate, and the Lisp-string sources of buffer-scoped sessions. classify_display_property takes the conditions and the object kind (string vs buffer first/last rule).

Under line numbers both :align-to resolvers hand the pixel calculator GNU's numbers (whole box width plus the field) and measure from the text-area edge, so a numeric target and center both move by the whole field as in 31.0.90 (src/xdisp.c:30438-30441, 30493). Emacs master's half-field center is declared in comments and tests, not ported.

Declared gaps (in code comments)

• A FORM is evaluated once per equal form with its first occurrence's bindings; GNU evaluates every occurrence.
• The span end is the fontification budget, so boundaries past the last displayed row are evaluated where GNU would not reach them.
• GNU's single-winner rule for overlay properties (src/textprop.c:656-675) is not applied; every covering overlay's form is evaluated.
• Prefixes are read at the overlay start, GNU reads them at each row start (src/xdisp.c:25131-25132, 25180).
• Frame-local chrome/margin strings and readers without a snapshot keep the structural rule (non-nil holds), as does a FORM the scan did not see.
• No inhibit-eval-during-redisplay; dsafe_eval_handler's "Error during redisplay" line, the nested-redisplay abort and backtrace-on-redisplay-error are not ported.
• Hscroll/continuation pen terms and GNU's one-pixel stretch past the target (src/xdisp.c:32882-32883) are declared in the geometry path.

Verification

• Tests, red first, for the evaluator bindings and error rule, the scan (text props, overlays, window-scoped overlays, hidden-overlay strings, point-max anchors, buffer-local prefixes), the classifier rules, the row builder (center with and without a fringe), and a realized GUI frame selecting the graphic clause when the text-terminal clause is listed first.
• Rendered-surface readback on macOS: the centered dashboard image lands at the GNU x-coordinate at 1x, and where GNU's arithmetic puts it for the rendered image at 2x.
• Six adversarial review rounds; every confirmed finding is fixed in the follow-up commits (rooting across Lisp, Flow propagation, the 31.0.90 vs master center arithmetic, overlay-string placement through the walk's resolver, after-string buffer-position at the overlay end, point-max anchors).
• Linux and Windows not run; the branch has no platform-specific code.

Found while comparing with GNU but not in this PR: at 2x images auto-scale from the device column width and render smaller than GNU's, and the native macOS menu bar is not implemented (menus render in-frame on every platform).

🤖 Generated with Claude Code

https://claude.ai/code/session_01NxL1hNkQvGiUyXAHJJuyXC

[tag-und-nacht]

Reviewed at 86e193634108910704ccfd09aa38a0fb77d8287c against the commit history, conversation export, and GNU Emacs 31.0.90 source. I found four issues to fix before merging:

1. [P1] Changed conditions do not invalidate retained rows — engine.rs:3554

   The engine chooses row reuse before evaluating conditions, then reuses those rows even when a condition changes. Reproduced: a true condition displays "REPLACED"; toggling it to nil evaluates the condition again but leaves "REPLACED" visible. A fresh engine correctly displays the underlying "X". The probe reported cursor_only_windows: 1 and reused_rows: 2. Condition results need to participate in reuse validity.

2. [P2] A disabled occurrence suppresses an enabled occurrence elsewhere — display_when/mod.rs:86

   Deduplication checks only the form, before checking eval_enabled. If (disable-eval (when (= 1 1) . "X")) precedes an ordinary (when (= 1 1) . "Y"), the disabled occurrence caches false for both. A layout probe using these properties on the two lines of a\nb\n displayed a and b, instead of a and Y. GNU applies the disabled-evaluation policy separately to each specification (xdisp.c:6139–6140). Disabled evaluation must remain specific to its occurrence.

3. [P2] The scan executes unreachable string clauses — display_when.rs:270

   For a prefix whose display property is ("X" (when (progn (setq probe t) nil) . "Y")), GNU stops after the first replacement and never evaluates the condition (xdisp.c:6034–6040). This scan executes it anyway: a layout probe confirmed that probe becomes true. The classifier’s later short-circuit cannot undo these side effects; evaluation must honor the same stopping rule.

4. [P2] Conditions see the wrong point in nonselected windows — display_when/mod.rs:45

   Evaluation selects the window’s buffer but does not install its point. With two windows sharing a buffer, selected-window point 1 and other-window point 5, (when (= (point) 5) . "YES") incorrectly fails in the other window. The probe displayed abcdef there instead of YESbcdef. GNU temporarily installs that window’s point during redisplay (xdisp.c:20578 onward). The evaluation context needs equivalent setup and restoration.

Validation: 39 existing scoped tests passed, including alignment and conditional-prefix integration tests. Four additional temporary regression probes reproduced the findings above; all source edits were restored after the review. The full suite and GUI pixel checks were not rerun for this review.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.