pmacs/docs/bottom-panel-stage2-framing.md

38 KiB
Raw Blame History

Bottom panel Stage 2 — the GPU panel band (framing)

Revision 4 — pre-implementation. Ground truth: canonical main @ ccf29e3, protocol v20, 2026-07-25.

Stage 1 (#155, merge e745068) gave pmacs window placement, window parameters, TUI side windows, the divider, and the adopter display opt-in. It deliberately set FrontendView::panel_capable = false for every semantic session, so a GPU frontend silently falls back to the non-side target. Stage 2 flips that bit, under an exact negotiated rule, and earns the right to.

This document is the re-framing docs/bottom-panel-framing.md (rev 4) §2 requires before Stage 2 is implemented. It does not restate the parent's decisions or replace its acceptance criteria. It records the re-scout against current main, closes the four scout obligations review round 1 required, and fixes what round 1 found wrong.

Inherited reading, all of which remains authoritative: parent Q#BP8 (the band), Q#BP9 (protocol), Q#BP14 (the primary-document projection contract and its census classification), Q#BP14a (panel input gating is per-window), Q#BP14b (focus chrome and per-window overlay routing), Q#BP15 (PanelFrame lifecycle), Q#BP15a (three geometries), Q#BP16 (pointer transport), Q#BP17 (fold projection), and parent acceptance criteria 3755.

0. Revision history

0.0 Round 3 (rev 3 → rev 4) — 1 blocking, 1 high, 1 medium, all closed

  • R3-1 (blocker). Rev 3's three-boundary model was right but its call-site table was wrong in five places, and each error was a real defect: :6140 is document completion placement (classified status-owned, which would let completion overlap the panel); :7195/:7212 are the two status text bounds (classified document-owned); :7351 clips global minibuffer candidate glyphs to the dropdown's band anchor (classified document-owned, which would clip them against the document boundary); :8561 (document edge scrolling) was missing entirely, leaving it tied to the old bottom; and :8077 was described as completion placement when it is caret clipping (its class was right, its label wrong). §5.3's table is rebuilt from the full census and every row is verified against the source. Root cause worth recording: rev 3's table was built from a grep | head -20 over 29 matches. The truncation is exactly why :8561 vanished. The census is now stated as 20 production sites + 1 definition + 8 test sites = 29, so a future reader can check the arithmetic instead of trusting the list.
  • R3-2 (high). The three equations permitted negative coordinates on a surface shorter than its chrome, where today's text_area_bottom clamps with .max(0.0). All three are now explicitly clamped, preserving the current helper's behavior.
  • R3-3 (medium). §5.1's "exact split" omitted validate_cells's cell.attachment.is_some() rejection. It is now classified — and shared, with the reasoning pinned.

0.1 Round 2 (rev 2 → rev 3) — 1 blocking, 2 high, 1 medium, all closed

  • R2-1 (blocker). Rev 2's "one document-bottom seam" conflated two boundaries that must diverge once a panel exists. Several sites it named are not document-bottom consumers at all: the status-band background (main.rs:5908) must stay at the physical window bottom, the status text buffers (:3175, :3185, :6601, :6607) consume a height and never a bottom coordinate, and status text placement (:7134) sits inside an unchanged band. §5.3 now splits the single value into three named boundaries, classifies every existing text_area_bottom call site, and adds the contrast assertion that catches a uniformly-wrong implementation moving both together.
  • R2-2 (high). accept_frame_geometry -> bool cannot distinguish advanced from accepted duplicate from rejected. It now returns an explicit three-valued result. The exhaustion wording also permitted retaining stale geometry, which is not fail-closed: §3.1 now clears the authoritative declaration and reconciles to hidden, and adds the frontend-side terminal latch.
  • R2-3 (high). Parent acceptance 52 was assigned wholly to 2A, but 2A has no semantic panel projection — it can only prove the extracted painter accepts an explicit None. 52 is now also reasserted in 2B, where the contract becomes production-reachable.
  • R2-4 (medium). §9 names the four touched acceptance suites explicitly rather than relying on "standing suite".
  • Both §8 open items are decided (§5.3): BASE_DIVIDER_HEIGHT = 4.0 at scale 1.0, and TEXT_TOP stays unscaled.

0.2 Round 1 (rev 1 → rev 2) — 2 blocking, 3 high, 3 revision points, all closed

  • R1-1 (blocker). Rev 1 said all 23 census reads route through primary_document_window. That contradicts Q#BP14, which routes only the Projection class (#1#12, #21#22) that way and leaves focus, input, chrome, and bell consumers on their own authorities. Rev 1's rule would have broken remote-op validation, DispatchIdle, presence, focused search/menu/completion routing, and bell ownership. §3.2 now restores all four classes; §7's criterion pins them separately. The inherited-reading list above gains Q#BP14 and Q#BP14a.
  • R1-2 (blocker). Rev 1 treated the three src/statusline.rs active reads as one disposition. Only :644 selects the wrong window; :629 and :675 must keep tracking actual focus. §3.3 is rewritten and the criterion states the required behavior instead of routing focus away.
  • R1-3 (high). The panel_capable flip needed an exact attach rule, not "for semantic sessions". §3.5 states it: v21-or-later negotiated authenticated semantic session only.
  • R1-4 (high). Option 1 accepted, but the epoch needed a state machine, split APIs, and a fail-closed allocator. §3.1 now carries the transition table and the API split. Rev 1's phrasing "rejects a lower-or-equal epoch carrying different data" was itself wrong — a lower epoch carrying identical data is still stale.
  • R1-5 (high). Rev 1's eleven draft criteria silently omitted parent 3755. §7 now declares the parent list authoritative, maps it to 2A/2B, and adds only refinements. The painter-extraction criterion pins cursor, view_top, and passive-window state, not just cells.
  • R1-6. All four scout obligations are closed in §5.
  • R1-7. The coherence statement understated journey impact and overclaimed on background work. §6 names journey steps 710 and narrows the §9 claim.
  • R1-8. Factual corrections in §1 and §3.2.

1. Anchor re-scout

Parent anchor Now at Verdict
paint_frame returns cursor separately (editor.rs:2833) src/editor.rs:3171 Holds
Cursor-visible prep (editor.rs:2883-2935) src/editor.rs:3249+ Holds; Stage 1 inserted work above it (§2)
Per-window paint body (editor.rs:2937-3040) after src/editor.rs:3260 Holds
fold_map_for_window gates on the active frontend (editor_core.rs:566) src/editor_core.rs:734, gate at :738 Holds
Stale "semantic session never enters paint_frame" (window.rs:339) src/window.rs:562 Holds, still stale; now embedded in a longer fold_projection doc block, so the edit is a paragraph rewrite
Mouse is contractually the grid path (daemon.rs:3122-3130) src/daemon.rs:3123 Holds
Permanent 24×80 placeholder (attach.rs:420-429, :573-577) pmacs-gpu/src/attach.rs:577, single site Holds
Byte pin InstanceMessage::InitialTargetResult pmacs-protocol/src/message.rs:1145 Holds — still the enum's final variant
Byte pin FrontendEvent::TerminalPointer final variant of its enum Holds

Protocol is still v20 (pmacs-protocol/src/message.rs:1568); no intervening PR bumped it. Q#BP9's conditional resolves: Stage 2 is v21, no reservation was taken and none was needed.

Fifteen PRs merged between the parent's last re-scout (47581f4) and this one: #149, #150, #152#155, #158#166. Nothing in the parent's mechanical model was falsified by any of them.

2. What Stage 1 already built for Stage 2

  • DeclaredFrameGeometry { geometry_epoch: u64, total: CellSize } (src/window.rs:522-528), held as FrontendView::frame_geometry: Option<_> (:589) where None means unknown — Q#BP15a's "unknown is first-class", already landed.
  • EditorState::sync_frame_geometry (src/editor.rs:877-882) → declare_frame_geometry + reconcile_panel_layout, driven from two daemon sites gated on panel_capable_for (src/daemon.rs:1882-1883 attach, :1972-1973 resize).
  • paint_frame declares geometry itself (src/editor.rs:3187), before the statusline fan-out and before the long mutable core borrow.
  • StatuslineEvaluationTarget (src/statusline.rs:212-226) is already a two-variant enum, so Q#BP8's fan-out generalization is an added variant, not a refactor.
  • primary_document_window (src/editor_core.rs:2830) and primary_document_buffer (:2845).

3. Findings and decisions

3.1 Q#BP2S1 — epoch ownership, resolved: frontend-owned, with an exact state machine

Decision: option 1. The epoch is owned by the frontend for negotiated semantic-panel sessions. The deciding argument is one rev 1 missed: a font or scale transaction can require invalidating an old PanelFrame even when the derived CellSize is identical. Daemon value dedup cannot detect that case, because the cell totals it compares are unchanged while the pixels behind them are not.

The landed allocator conflicts in three ways (src/editor_core.rs:3155-3172): it allocates the id itself, it early-returns when total is unchanged (value dedup), and it uses saturating_add, which is neither wrapping nor fail-closed — it pins at u64::MAX, after which two different geometries share one id.

Acceptance rules for a semantic declaration:

Incoming declaration Result
epoch greater than stored Accept, store verbatim, even if total is unchanged
same epoch, same total Idempotent no-op
same epoch, different total Reject
lower epoch, any total Reject

The last row is deliberate and corrects rev 1: a lower epoch carrying identical data is still stale and must not be accepted.

API split. Two methods, not one method with an optional epoch:

  • declare_frame_geometry(fid, total) — the grid/LOCAL allocator. Keeps value dedup (correct there: cells are the unit, and an unchanged grid means an old frame is still valid under unchanged metrics). Changes from saturating_add to checked allocation with an explicit fail-closed exhaustion arm.
  • accept_frame_geometry(fid, geometry_epoch, total) -> GeometryUpdate — the semantic path. No value dedup; applies the table above verbatim.

An ambiguous single method with an Option<u64> epoch is rejected explicitly: it would let a future caller silently take the wrong regime.

The result is three-valued, not a boolean. A boolean cannot distinguish the three outcomes the caller must act on differently:

enum GeometryUpdate {
    /// Epoch advanced: stored verbatim. Run panel reconciliation.
    Advanced,
    /// Same epoch, same total: already current. Do no work.
    Duplicate,
    /// Same epoch with different total, or a lower epoch: stale or
    /// conflicting. Drop the event before any reconciliation.
    Rejected,
}

Advanced reconciles, Duplicate returns without touching panel state, and Rejected drops the event. Collapsing Duplicate into either neighbour is a defect in one direction or the other: folded into Advanced it reconciles on every repeated declaration, folded into Rejected it would log or surface a stale-event condition that never happened. (If a boolean is kept for a narrower internal caller, it must be named advanced, never acceptedDuplicate is accepted.)

Initial epoch. The frontend's first declaration after attach acceptance carries epoch 1. 0 is reserved as "never declared" and is rejected on the wire.

Exhaustion fails closed on both sides, and rev 2's wording did not. Saying the panel "stays at its last valid geometry" is not fail-closed: if the real frame resizes after the allocator is exhausted, the daemon would keep painting a panel sized to geometry that no longer describes the frontend.

  • Grid/LOCAL path. On checked-allocation exhaustion, clear the authoritative frame_geometry (back to None = unknown) and reconcile. Unknown is already non-presentable under Q#BP2b, so the panel hides. Stale geometry is never retained.
  • Frontend path. On exhaustion the frontend sets a terminal latch for the life of the session: it sends no further geometry, and — critically — an old matching Present cannot make the band reappear, because the latch suppresses paint and hit-testing independently of frame validity. Only a fresh session (reconnect) clears it. Without the latch, a retained Present whose epoch still matches the last declaration would resurrect a band under geometry the frontend has disowned.

3.2 The census is classified, and it is mostly unrouted

Correction to rev 1. Q#BP14 routes only the Projection class through primary_document_window. Rev 1's "all 23 reads" was wrong and would have broken five subsystems. The four classes, restored:

Class Census items Authority
Projection #1#7, #9, #10, #12, #21, #22 primary_document_window / primary_document_buffer
Projection + focus #8 (document Pointer), #11 (full-window TerminalPointer) Align the primary document window and then activate it — the one place the two legitimately move together
Focus / input #13 (remote-op validation), #14 (dispatch_idle_for), #15 (presence), #23 (remote-op application) The frontend's actually focused window. Q#BP14a: gating is per-window, never per-buffer
Focus chrome / surface-routed #16#19 (search, menu, minibuffer, completion) Q#BP14b's routing table — the currently owned surface, with authoritative clears for the other
Focus / session #20 (terminal bell drain) Per-session counter; the focused window chooses which session may drain

Rerouting any of the last three classes to the document is a defect, not a simplification: it would break remote-op validation and application, DispatchIdle, presence, focused search/menu/completion routing, and bell ownership.

How much is already routed. primary_document_window has four references in src/ and two production paths: directly at src/daemon.rs:1639 (#148's initial-target bootstrap, Q#BP11b), and through primary_document_buffer at src/daemon.rs:2998, which is census #22 and carries a comment naming it. So one census item is routed and the Projection class is otherwise open. For scale, src/*.rs still holds ~80 non-test direct .active reads on top of the active_window* / active_buffer* helper family (src/editor_core.rs:663-967).

This is not a Stage 1 defect — with panel_capable = false no semantic frontend can hold a side window, so the unrouted Projection reads are unreachable from the GPU. It does mean classified census routing is the bulk of Stage 2, which is why it is Stage 2A.

3.3 The three statusline reads have two dispositions, not one

All three sites are real, but only one is wrong:

  • src/statusline.rs:644.get(&view.active) selects the wrong window when a panel is focused. This is the Projection read (#12).
  • src/statusline.rs:629 and :675active: window_id == view.active must continue tracking actual focus. Three reasons: grid contexts need a truthful active; post-callback revalidation must notice a focus change; and parent acceptance 42 explicitly requires that a document provider may observe active = false while the panel is focused.

The new semantic-layout target therefore captures the primary document window plus the visible side window, marks each context active iff its window_id == view.active, invokes each provider exactly once, and invalidates the entire evaluation if a callback mutates layout or focus. Unprojected document splits run no callbacks (Q#BP8). Route the primary-document result to semantic StatuslineSegments and the side result to the panel mode line.

3.4 Fold projection

Unchanged from Q#BP17, with the anchor corrected: the extracted painter takes the map as a parameter; the panel path passes None when the owning frontend's fold_projection is false and must never call fold_map_for_window, which gates on the active frontend (src/editor_core.rs:734, gate at :738) — right for command-time reckoning, wrong for painting another frontend's panel. The stale comment is at src/window.rs:562.

3.5 The panel_capable flip needs a negotiated rule

Not "true for semantic sessions". Exactly:

panel_capable = true only for an authenticated semantic session that negotiated v21 or later.

A v6v20 semantic frontend stays non-panel-capable and takes the existing Stage 1 fallback: the non-side target with every side-specific parameter discarded, leaving the document window undedicated (Q#BP2c). "It receives no new events" is insufficient — if the daemon nevertheless places that frontend's window in a side panel it cannot render, the window becomes invisible. The gate is on placement, not only on transport. Parent acceptance 51 pins the mixed session.

4. Revisions to the parent framing

Only these; everything else stands.

  • Q#BP9 resolves to v21.
  • Q#BP15a's epoch ownership is specified by §3.1's table and API split, replacing the parent's one-line "frontend-owned" statement.
  • Q#BP8's statusline criterion splits per §3.3: one read reroutes, two keep tracking focus.
  • Q#BP17's stale comment is at src/window.rs:562, and parent acceptance 52's reference to :339 should be read against that.

5. The four scout obligations, closed

5.1 The shared cell-grid validator boundary

TerminalFrame::validate (pmacs-protocol/src/terminal.rs:226) currently interleaves both concerns. The exact split:

  • Factored into the shared parameterized wire-cell-grid validator: checked area (the checked_mul + usize::try_from guard), the MAX_TERMINAL_VISIBLE_CELLS = 262,144 aggregate cap, cell-count equality against declared area, cursor-in-bounds, and validate_cells's glyph width / continuation topology and aggregate glyph-byte checks.
  • Stays terminal-only: the MAX_TERMINAL_ROWS/COLS = 512 per-axis caps in checked_area, validate_metadata for title/signal/crash text, validate_selection, and the at_bottom == (scroll_offset == 0) coupling.

Attachment rejection is shared, not terminal-only. validate_cells also rejects cell.attachment.is_some() (pmacs-protocol/src/terminal.rs:305), and its error text reads "A cell carries a frontend attachment, which terminals never use" (:190-191) — phrased as a terminal-specific fact, which is why rev 3 missed it. Stage 2 classifies it shared: panels implement no attachment rendering, so a PanelFrame carrying one describes a surface the GPU would silently not draw. Shared rejection fails closed on the producer side rather than shipping an invisible cell. The error message is reworded away from "which terminals never use" to a grid-neutral phrasing when it moves. If a later stage gives panels attachment rendering, this rejection moves back to terminal-only as a deliberate, reviewed change — not by default.

PanelFrame takes the shared half plus its own presence/epoch rules and does not inherit the 512 per-axis cap (Bet B5'), so a 4K small-font panel wider than 512 columns is legal while the shared area budget still binds. Parent acceptance 39 pins exactly this.

5.2 The GPU outbox needs four more tags

coalesce_kind (pmacs-gpu/src/attach.rs:331) today returns four tail-only tags: Viewport → 0, Pointer{Drag} → 1, TerminalPointer{Move} → 2, TerminalPointer{Drag} → 3. Everything else is None = lossless, counting against OUTBOX_MAX = 8192.

Stage 2 adds four distinct tags: FrontendCellGeometry → 4, PanelResizeRows → 5, PanelPointer{Move} → 6, PanelPointer{Drag} → 7. Geometry is latest-wins (epochs need only increase, not be consecutive); resize drag is latest-wins over the complete event including its epochs. PanelPointer Down/Up/wheel/context stay lossless and ordered — repeated left Downs are what the daemon click state reads as a multi-click, and Down(Right) is the context-menu gesture. Tail-only replacement preserves ordering across an intervening event of any other class.

5.3 The pixel formula's inputs — and one trap

The formula in Q#BP15a is contract-level, not an implementation detail, because its inputs are not all safe to adopt:

Input Source Note
status_band_height_px FontMetrics::status_band_height (pmacs-gpu/src/main.rs:137) = BASE_STATUS_BAND_HEIGHT * scale Safe
TEXT_TOP_px const TEXT_TOP: f32 = 16.0 (main.rs:352) Safe; unscaled today
code_line_height_px FontMetrics::code_line_height (main.rs:131) = BASE_CODE_LINE_HEIGHT * scale Safe
resolved_monospace_advance_px State::mono_advance (main.rs:4899) Unsafe to adopt blindly
divider_height_px BASE_DIVIDER_HEIGHT Does not exist yet

The mono_advance trap. State::mono_advance returns measured_mono_advance when a FontFacts probe has been applied, but otherwise falls back to the first shaped glyph of the document buffer (main.rs:4903+). Panel column count would therefore become document-dependent: two GPU frontends showing different files could derive different total.cols from identical metrics, and the same frontend's panel width could change when the document's first glyph changes.

Decision. The panel geometry declaration uses a stable normal-face probe, never the document sample. probe_mono_advance(font_system, family, metrics) (main.rs:323) already exists and is exactly this: it shapes ADVANCE_PROBE in a scratch buffer, independent of document contents, dividing total run width by logical cells so ligature substitution survives. The declaration resolves its advance from that probe for the current family/metrics. If the probe returns None (the family shapes no width), the frontend declares zero usable geometry under a new epoch — the panel hides — rather than falling back to a document sample.

BASE_DIVIDER_HEIGHT = 4.0 at scale 1.0, scaled by FontMetrics::scale like status_band_height. A 12 px rule is adequate decoration but too fragile as the drag hit strip; 4 px still reads as a rule while giving the pointer a usable target. The entire strip is painted with ui.divider, and that exact rectangle is the hover/drag hit region — paint geometry and hit geometry are the same rect, so they cannot drift apart.

TEXT_TOP stays 16.0, unscaled. It is a fixed surface inset today, like TEXT_LEFT and the other paddings, while FontMetrics::scale governs font-derived metrics and row chrome. Scaling it only inside the declaration formula would disagree with the actual renderer; scaling every renderer and hit-test occurrence is a wholesale inset/DPI change and is named here as separate work, not smuggled into Stage 2. The formula is pinned to the real unscaled inset.

Accordingly, Q#BP15a's "all quantities use the frontend's current scale" is narrowed: font-derived metrics and the divider scale; fixed surface insets keep their current units.

The seam is three boundaries, not one

Rev 2 asked for a single document-bottom accessor. That was wrong: once a panel is installed, today's single value must diverge into three, because some of its consumers must not move at all.

All three clamp at zero, preserving today's text_area_bottom .max(0.0) behavior — without the clamps a surface shorter than its own chrome yields negative coordinates, and the "exact formula" stops being exact precisely where it matters most:

status_band_top          = max(0, surface_height - status_band_height)

geometry_capacity_bottom = max(0, status_band_top - divider_height)
                           // divider reserved even while absent

document_text_bottom     = max(0, status_band_top
                                  - installed_panel_height
                                  - installed_divider_height)

geometry_capacity_bottom is what Q#BP15a's asymmetry already requires: the divider is subtracted for sizing purposes even while the panel is absent, which is what breaks the first-open cycle, while the document renderer does not actually lose those pixels until a Present panel is painted.

Today text_area_bottom (pmacs-gpu/src/main.rs:8490) is all three at once, and its doc comment calls it "the single source for every bottom-of-text computation" (Q#S3).

The census is 29 matches: 20 production call sites, 1 definition (:8490), and 8 test sites (:12887, :12937, :12997, :13109, :13793, :14013, :15306, :15386). Every production site, classified individually against the source:

Status-owned — must stay pixel-identical at the physical window bottom, using status_band_top (8 sites):

Site What it is
:5908 Status-band background rect y
:6003 mb_visible_window — rows that fit above the band
:6027 mb_dropdown_window origin — dropdown grows up from the band
:7134 status_top for the right status group
:7195 status_buffer TextBounds.top — status text bound
:7212 status_left_buffer TextBounds.top — status text bound
:7351 Minibuffer candidate glyph clip, anchored to the dropdown's band origin
:7922 status_top, second site

The minibuffer is global, bufferless chrome anchored to the status band (Q#BP14b keeps MinibufferPrompt global), so all four of its sites — :6003, :6027, :7351, and its status_left_buffer bound :7212 — stay status-owned. Clipping candidate glyphs at document_text_bottom would clip the dropdown against a boundary it does not sit above.

Document-owned — must move when a band is installed, using document_text_bottom (12 sites):

Site What it is
:4566 terminal_cell_viewport — drawable height for the cell grid
:6118 completion_anchor_px — anchor visibility bottom
:6140 completion_dropdown_layoutdocument completion placement; band_top - (line_top + line_h) is the space below the anchor line
:6581 code_height
:7174 Code text clip bottom
:7242 Math text clip bottom
:7273 Gutter clip bottom
:7421 Terminal clip bottom
:8077 code_caret_rect_in_clipcaret clipping
:8497 Minimap drawable height
:8501 Visible-line estimate
:8561 edge_scroll_directiondocument edge scrolling

Geometry declaration uses geometry_capacity_bottom, and is the Q#BP15a conversion only.

Sites that consume no bottom coordinate at all and must not be touched: :3175, :3185, :6601, :6607 size the status text buffers to status_band_height directly. Rev 2 listed them as seam consumers; they are not.

Three of these classifications are the ones a plausible implementation gets wrong, and each has a visible symptom: document completion (:6140) anchored to status_band_top overlaps the panel; minibuffer candidates (:7351) clipped at document_text_bottom are cut off; and edge scrolling (:8561) left on the old bottom auto-scrolls from inside the panel.

Each call site is classified individually. A blanket rewrite of text_area_bottom to subtract the band would move the status chrome with the document and is the defect this section exists to prevent.

The contrast assertion (A2B-4). "Every document consumer moved" is only half a test — a uniformly wrong implementation that moves everything passes it. The criterion must assert both directions in one scenario: installing a panel moves every document-owned consumer while the status band stays pixel-identical at the physical window bottom. That is the assertion a blanket rewrite fails.

The one-accessor-per-boundary rule still holds within each class: a second, unrouted derivation of any of the three is the exact shape of Stage 1's Layout::compute two-caller defect, where src/overlay_paint.rs derived its own rect and painted peer cursors at unfixed rows.

5.4 Ordering against folding Stage 3

Settled by review round 1: bottom-panel Stage 2 first, through the landed GPU band. Folding Stage 3 then re-scouts the extracted painter, the panel projection, clipping, and fold_projection behavior exactly once.

6. Coherence impact (per COHERENCE.md §20)

  • Journey steps touched: four, on the GPU frontend — steps 710 (find symbol / find file, terminal, build and test, error inspection). Rev 1 said "none directly", which contradicted its own next sentence. Today a GPU user who triggers references, project search, a terminal, compile, or error inspection gets the Stage 1 non-side fallback: the output surface steals a document window instead of opening a panel. Every one of those steps therefore behaves differently on GPU than on TUI, and Stage 2 is what closes the divergence.
  • Interaction islands added: none, and this is a reduction. §6 grades islands "weak, and growing by one island per modal feature". Stage 2 extends one already-adopted policy (display = "panel", used by listview, compile, and terminal) to a second frontend rather than minting a GPU-only surface. Q#BP14b deliberately reuses the existing SearchPrompt / MenuPrompt / CompletionPopup messages instead of panel-specific twins.
  • Config registry adoption: inherited, not extended. Stage 1's window.panel-height and window.min-height already live in the registry. Stage 2 adds no new user-facing option; if the band needs one, it enters the registry.
  • Background-work attribution: unchanged, and this stage does not advance it. Rev 1 implied Stage 2 helps §9's activity-view gap. It does not. A panel gives output a coherent placement; it does not make terminal PTYs, LSP servers, or workers appear in the activity/ownership view §9 describes, and it adds no join key across the four disjoint activity planes. The §9 gap is untouched.
  • Section this serves: COHERENCE.md §14, which records the panel primitive as landed for Stage 1 and names "Stage 2 (GPU band) pending its own framing" as the open item.

7. Acceptance

Parent criteria 3755 remain authoritative and are not replaced. This section maps them to the two slices and adds only refinements.

7.1 Stage 2A — classified census routing + painter extraction

No protocol change. Parent criteria that apply in full: 42, 43, 44, 51 (the LOCAL-panel inheritance half), plus the extraction half of 52.

52 splits across the slices. 2A has no semantic panel projection and no PanelFrame, so all it can prove is that the extracted painter honors an explicitly supplied None fold map and that the stale src/window.rs:562 comment is corrected. The actual contract — a semantic panel with fold_projection = false never collapses folds and never calls fold_map_for_window — is production-reachable only once 2B lands the projection and the capability flip. It is therefore reasserted in 2B (§7.2).

Refinements 2A adds:

  • A2A-1 (replaces rev 1's criterion 1). Every Projection census item (#1#7, #9, #10, #12, #21, #22) resolves through primary_document_window / primary_document_buffer; #8 and #11 align and then activate; #13, #14, #15, #23 continue to resolve the actually focused window; #16#19 follow Q#BP14b's routing table; #20 keeps its per-session counter with focus choosing the eligible terminal. Each class is asserted separately, at the outermost user-reachable seam, and falsified by revert. A test that only proves "the document is used" would pass with the focus classes wrongly rerouted, so the focus-class assertions are the load-bearing half.
  • A2A-2 (replaces rev 1's criterion 2). src/statusline.rs:644 resolves the primary document window, while :629 and :675 continue to report actual focus — pinned by a document provider truthfully observing active = false while the panel is focused (parent 42). The semantic-layout target captures primary document + visible side window, invokes each provider exactly once, and invalidates the whole evaluation when a callback mutates layout or focus.
  • A2A-3 (replaces rev 1's criterion 3). The painter extraction preserves, for grid frontends: the painted cells, the returned cursor, the focused window's view_top mutation from the auto-scroll clamp, and passive windows' untouched view_top and scroll state. Byte-identical cells alone would not catch a clamp that silently moved to the wrong window.

7.2 Stage 2B — v21 protocol + daemon projection + GPU band

Parent criteria that apply in full: 37, 38, 39, 40, 41, 45, 46, 47, 48, 49, 50, 51, 53, 54, 55, plus re-assertion of 42, 43, 44, and 52 through the actual negotiated capability flip rather than through a test-only panel-capable semantic view. 52's 2B form is the production one: a real semantic frontend with fold_projection = false displaying a folded buffer in a panel shows every source line, and the panel path never reaches fold_map_for_window.

Refinements 2B adds:

  • A2B-1. The epoch state machine of §3.1 is pinned row by row, including the lower-epoch-identical-data rejection and the same-epoch-different-total rejection, and each row's Advanced/Duplicate/Rejected result is asserted — a Duplicate performs no reconciliation and a Rejected mutates nothing. Epoch 0 is rejected on the wire. Exhaustion is pinned on both sides: grid exhaustion clears frame_geometry to unknown and the panel hides (a subsequent real resize must not paint a stale-geometry panel), and a frontend that exhausts latches — a retained Present whose epoch still matches cannot make the band reappear, and only a fresh session clears the latch.
  • A2B-2. A font or scale change that leaves CellSize identical still produces a new geometry_epoch, and the older PanelFrame neither paints nor hit-tests until a matching Present arrives. This is the case daemon value dedup cannot see and is why option 1 was chosen.
  • A2B-3. Panel columns are derived from the stable normal-face probe, not State::mono_advance's document-glyph fallback: two GPU frontends with identical metrics and different documents derive identical total.cols, and a probe returning None declares zero usable geometry rather than falling back to a document sample.
  • A2B-4 (contrast assertion). Installing a panel moves all twelve document-owned consumers of §5.3 by exactly installed_panel_height + divider_height, while all eight status-owned sites stay pixel-identical at the physical window bottom. Both halves are asserted in one scenario: a uniformly wrong implementation that moves the status band too passes the "everything moved" half alone. Three rows carry their own named symptom because they are the ones a plausible implementation misclassifies — document completion (:6140) must not overlap the band, minibuffer candidates (:7351) must not be clipped by it, and edge scrolling (:8561) must not trigger from inside it. The geometry declaration separately reserves the divider while the panel is Absent, and the document loses no pixels until a Present is painted. All three boundaries clamp at zero on a surface shorter than its chrome.
  • A2B-5. panel_capable is true only for a v21+ negotiated authenticated semantic session; a v20 semantic session is never placed in a side window, not merely denied the events.

8. Open items

None. Both round-1 open items are decided in §5.3: BASE_DIVIDER_HEIGHT = 4.0 at scale 1.0 (scaled, whole strip painted ui.divider and used as the hit rect), and TEXT_TOP stays unscaled with wholesale inset/DPI scaling named as separate work.

One deferral is recorded rather than resolved: wholesale surface-inset scaling (TEXT_TOP, TEXT_LEFT, and the sibling paddings under FontMetrics::scale) is pre-existing behavior Stage 2 pins rather than fixes. It belongs to a spacing-system change of its own.

9. Slices, branches, and gates

Per review round 1: two serial implementation PRs, each a named slice under this framing so one-feature/one-branch/one-PR holds. 2A lands before 2B branches — not stacked.

  • Stage 2A — classified census routing + per-window painter extraction. Branch bottom-panel-stage2a. No protocol change. The three-boundary GPU split is 2B, not 2A: it is only observable once a band can be installed.
  • Stage 2B — v21 protocol, daemon panel projection, GPU band, and the negotiated panel_capable flip. Branch bottom-panel-stage2b, cut from main after 2A merges. Repeats 2A's relevant census assertions through the real capability flip.

Gates for both: the standing suite from CLAUDE.md, plus the touched acceptance suites named explicitly — the standing rule is to run the suites a change touches, and "standing suite" does not name them:

  • bottom_panel_stage1_acceptance — the substrate both slices build on.
  • bottom_panel_stage2a_acceptance / bottom_panel_stage2b_acceptance — new, one per slice.
  • statusline_segments_acceptance — the fan-out target change (§3.3).
  • m11_5_semantic_acceptance — the semantic census (§3.2).
  • gpu_initial_target_acceptance — parent criterion 55.
  • gpu_font_acceptance — font/scale geometry refresh (§5.3), including the normal-face probe and the unscaled-TEXT_TOP decision.
  • The three vterm suites — the panel hosts terminals.
  • Folding Stage 2's 48 — shared projection.
  • PMACS_REQUIRE_GPU=1 cargo test -p pmacs-gpu.

Protocol round-trip and byte-pin tests ride 2B. Parent criterion 54's --headless-probe run — one real daemon, real PTY, real wgpu, through a panel-hosted terminal — is a 2B gate.