ARCADIA-Studio/harness/experiments/focus-optical-20261001/plan.md

Two-axis Focus optical experiment — predeclared plan

Status: experiment plan before implementation/native observation, 2026-10-01 PDT. Session 01a0d717/root owns this bounded part of issue 38, claim comment1437. The isolated branch is codex/focus-workspace-experiment-01a0d717-20261001, now based on published save-diagnostics handoff bc69c6d449641fc4395feae0d8373f62257c763c. PROJECT_GOAL.md is authoritative. This experiment cannot close dense-diagram usability or accept an editable workspace.

Question, alternatives and decision

Does recovering both width and height materially improve complete-scene reading at the same native window/DPI while retaining every model/diagram record and the current renderer? The prior restricted route generator, one-line association cue and E/F multiline label experiments remain distinct negative evidence, published on their own branches. Their code is excluded here.

A width-only trial is rejected for this question: checked historical routed bounds1506x814 and viewport826x534 predict only about16% Fit gain after both docks are hidden, because height then limits Fit. A hand-built blank-panel Baseline is rejected as the primary comparator because wrapping real chrome affects available height. Changing labels/routes, omitting relations, auto-layout or selected highlighting cannot answer this question and is excluded.

Choose an explicitly opted-in private, read-only Iced runner in the existing binary. Reuse DiagramCanvas::draw with fresh optical state; do not call its interactive update or production App::update/subscription. Freeze a validated Project and use App::with_project only as a presentation fixture for existing Baseline chrome. Baseline retains the actual body, task navigation, tool rows, docks, palette state and footer. Only the header action area may be replaced with same-height experiment controls; measure/document this deviation. Focus uses a compact visible header and status row with a full two-axis canvas. The renderer, projected/routed scene, labels, ports, authored metadata and model must remain identical. Normal launch keeps its existing application path.

Controls: Baseline/Focus, Fit and explicit measurement capture. No model editing, hover assistance, selection, manual camera, save/history/navigation integration is claimed. Inert chrome must exclude input before child dispatch, not merely drop resulting messages. Camera and measured size are privately owned and stamped per layout generation; reject stale, nonfinite, zero and repeated measurements. Every actual resize invalidates readiness, including resizes within one variant. A sensor receipt precedes the fitted draw: capture requires a later matching draw acknowledgment for variant/generation/viewport/camera. This proves CPU drawing was called, not GPU/display completion; the native screenshot verifies visible output. An unmeasured layout is visibly pending rather than using a guessed viewport.

Fixed fixtures and admission

Admit source bytes before launch and compare them again after normal exit. Every case runs against the complete unmodified Project, not a label-drag saved derivative. The optical runner rejects absent diagrams and all non-Architecture projections, including Scenario/StateMachine; compare complete Project after its Document construction, and preserve input bytes in the evidence. Errors/warnings are inventoried rather than silently repaired. The mixed case is source7; the two unchanged stock example files are source3. Use the existing persistence decoder/migration report for their declared admission. Freeze the complete decoded modern Project after admission, not the raw legacy JSON shape. This admission transformation is recorded separately from presentation; original source bytes remain untouched. Synchronize the chosen Architecture diagram/layer and derived Explorer/create-kind/palette presentation instead of only overwriting two App fields initialized for a different default diagram.

Case	Exact JSON / SHA-256	Saved Architecture diagram
Complete mixed case	eight-card.json, 138661bytes; 9db1015369f450d5810c96a01fddb8a009d57f6aafbda5182703f15ded23d9c8	00000000-0038-4000-8000-000000000001: eight visible cards,12 relations; actual projected ports/edge kinds recorded at admission
Unrelated nested OA	examples/routing-workshop.json, 100281bytes; 8592714f75b271133f21ce017d87896b6cf91ea0e4837f1c2ed453e21be26b14	945a7c88-a352-5028-b808-73d852e5abe6:12 authored placements; actual projected nested contents recorded
Unrelated nested SA	Same exact routing-workshop Project	cb21862d-79e1-5a9d-bdf0-ee344c2fb219:12 authored placements; actual projected nested contents recorded
Small ordinary view	examples/coastal-rescue.json, 49715bytes; 242f0b031229a87af5c4ebe5d651defe59c191a6b7a69dc0a77f7e203411f195	487c064d-df33-45f2-b4b7-6dd68d7b601f: OA functional dataflow,3 authored placements

These are existing examples plus the unchanged previously published representative case, not the user's unknown original eight-card model. They contain no named DataAssociation ends. Native end-label coverage is a recorded gap; it cannot be inferred from these cases. Do not claim broad projection, font, DPI or accessibility coverage. A later product decision must expand the fixture set appropriately.

Evaluators and stopping rules

1. Before native work, test the Fit formula against production margins/clamp, both limiting axes, offsets/empty scenes; stale/repeated sizing; and input exclusion using a spy widget. Mouse/button/wheel, keyboard/modifiers, touch, IME and window events other than allowed redraws cannot reach inert children. Draw-only canvas cannot publish actions, select, mutate gesture state or show interactive mouse feedback. No production reducer is reachable from the rig.
2. Native Windows comparison uses the same pinned binary/runtime, fixture, client size and observed DPI. Begin at Baseline Fit, capture settled measurements, Focus Fit and Baseline return. Record viewport, exact zoom/pan, routed bounds, scene counts, projection, palette/selection state, screenshot and process logs. Require complete full-scene containment at Fit, readable experiment controls, normal exit and unchanged inputs. Resizing must reject old measurements and update Fit without looping. Ordinary opt-out launch must still show its Home.
3. Exact whole Project, projected view, route point sequences/labels/end geometry and complete production SVG must be equal across Baseline/Focus/return and captures. Compare full values/bytes, not selected properties. Independently check Fit arithmetic from recorded viewport/bounds. A provisional0.85 Fit target on the complete mixed case is only an optical feasibility discriminator; magnification gain is not readability acceptance.
4. A reader without semantic answer-key/earlier-image exposure receives only the complete settled image and neutral questions: identify critical functional exchange endpoints/direction, the conversation participants and whether it has direction, ordinary versus port allocations, and port owners. Record exact answers, confidence, competing plausible ownership, illegible names and time before hover/selection/manual zoom. The key is withheld. Use Focus first for this new reading, while earlier Baseline first-reading results retain their original reader/build scope; this is not a randomized causal comparison. If no fresh reader is available, report that gap and do not claim a fresh pass.
5. A wrong critical association or unresolved competing ownership rejects Focus as a sufficient remedy even if scale improves. Larger text may remain a useful workspace hypothesis. Follow-up should address semantic distinction/shared stroke ownership rather than polish chrome or retune another label offset. An exposed parent's reading is assisted inspection only. A single agent is not a system-engineer study or general OA/SA acceptance.
6. Independent review covers source boundaries, invariants, evaluator adequacy, actual measurements/conservation, discrepancies and limitations. Archive useful results/failed probes with reproducible commands, exact source and raw evidence. No production promotion, issue closure, packaged release or editable-workspace claim follows from this experiment alone.

Editable promotion, if justified later, separately requires retained draft/caret/ widget/focus/input state, tools/held-gesture retirement, manual camera/Fit, pending-navigation and stale release barriers, Undo/Redo/Save/reopen and direct native review. Coordinate with peer-owned #43; no current ownership is reassigned. Linux/WSL remains user-deferred.

Independent proposal review required these admission/header/draw-ack refinements before native work. One fresh-reader spawn attempt failed with agent thread limit reached; this is a perception-evaluator availability gap, not a readability pass. Existing readers and parent know the answer key and cannot substitute for a fresh reader. Independent source/evidence review remains available.