Make the graph workspace readable across zoom and views

.agents/inbox/2026-10-28-graph-controls-need-shared-flow.md

@@ -0,0 +1,59 @@
+# Graph controls need shared flow and explicit event boundaries
+
+2026-10-28, PR #223 / issue #222. CI 40731080444 tested merge
+`7eb3b6085282c53b5a48ce806bd1e35bf21424c8` from source `b33febb`.
+Its durable-layout report contains four actual passes and five actual failures,
+plus the execution-failure record. All five failures show the connection Browse
+button intercepting pointer clicks intended for layout Retry/Discard controls.
+The independently verified smoke artifact SHA-256 is
+`c52d20e7` (prefix; complete digest retained in the GitHub artifact metadata).
+Screenshots and traces are retained in the failed run. Four downstream browser
+stages were skipped; this run is not acceptance.
+
+Cause: status and connection controls independently occupied bottom-left space
+at the same z-index. Increasing text size exposed an existing ownership problem.
+A second headed regression at 1440/390 opened the relationship browser's Edit
+button, but the parent document outside-click handler immediately cleared the
+selected edge, changing the new dialog into Create relationship.
+
+Corrective rule: put status, trace and connection controls in one responsive,
+scrollable flow dock in both item and group views. Include each control's actual
+footprint in routing/label obstacles. Do not fix overlapping controls by changing
+paint order or forcing test clicks. Keep browser-panel clicks inside its event
+boundary so opening an editor cannot also trigger outside dismissal.
+
+Validation commands use one headed Chrome, one worker and zero retries:
+- `playwright test item-wire-consumers.spec.ts layout-recovery.spec.ts --grep
+  'item wire paths|reload|unavailable storage'`: three passed in 36.7s. Exact
+  browser Edit now checks original source, target and type without data writes.
+- `playwright test layout-recovery.spec.ts --grep
+  'quota-failed|retry after real|pending layout keeps'`: three passed in 42.2s.
+  The pending-layout journey also checks disjoint status/Connections bounds,
+  normal Retry/Browse/Discard clicks at 1440 and 390px, and actual API/receipt
+  state. Existing durability and camera assertions remain unchanged.
+- Web `tsc --noEmit` passed using the owned dependency fixture. This is scoped
+  evidence, not a pinned clean-install or complete CI result.
+
+Limits: the dock reserves control space, not arbitrary card-to-overlay clearance.
+Full exact-source CI, actual artifact review and final dev checks remain required.
+The separate priority feature has additional desktop/phone browser coverage;
+its success does not establish this PR's full acceptance.
+
+The first phone trace run rejected the dock container as overlapping its own
+trace child. The transparent layout container no longer declares itself as a
+second obstacle; status, trace and connection controls each retain their actual
+measured obstacle. The oracle and its overlap assertions are unchanged. The
+failed `dock-trace-phone1` artifacts are retained.
+The corrected phone hierarchy/trace journey passed in 25.3s in
+`dock-trace-phone2`, including its original overlap/readability checks and the
+8px trace-text negative. No oracle assertion was relaxed.
+
+Independent review of `fe84973` used the actual dock/status components and the
+production D3 zoom filter. A wheel over status prevented native scrolling and
+changed camera scale from 1 to 0.8705505633. Marking the shared dock with the
+existing `data-lod-controls` boundary leaves the wheel default intact and scale
+at 1. The pending-layout browser case now adds a 390×500 viewport, requires
+actual dock overflow, wheels over its status area, and verifies native scrolling
+with an unchanged camera before discarding the owned receipt.
+The complete amended pending-layout journey passed in 15.7s on 2026-10-28,
+with one headed Chrome, one worker and zero retries (`dock-native-scroll1`).

.agents/inbox/2026-10-28-hidden-labels-need-valid-measurement.md

@@ -0,0 +1,87 @@
+# Keep conservative label geometry when hidden SVG measurement is empty
+
+2026-10-28 UTC, independent review of item-wire source
+`d9a0a5855612290b7c765999440b0d1d55b36333` for #220 / PR #221. Existing
+guidance was searched in the symptom index and the offset-label lesson. This
+was a source-review finding, not a reproduced deployed/browser incident.
+
+The first placement pass can hide an edge-label group before its delayed
+150ms sizing callback runs. Accepting an empty `getBBox()` then replaces a
+conservative text estimate with padding alone (36 units), allowing a long
+label to be admitted using an undersized obstacle/background. Initialize text,
+icon and background from one conservative cache, and replace that cache only
+when both measured dimensions are positive. Hidden labels must not become
+small merely because they are not currently measurable.
+
+The same review found a line-only diagnostic reading `x1.baseVal` after the
+renderer changed to SVG paths. Read actual path endpoints through
+`getPointAtLength` and transform them through `getScreenCTM`. Searching for
+remaining `SVGLineElement` / `x1.baseVal` consumers found this one diagnostic.
+
+A separate source audit found that directed perpendicular offsets put reverse
+edges on the same lane: reversing both the direction and lane sign cancels the
+separation. Use a canonical unordered endpoint direction to define lane sides,
+while preserving the original route direction for arrows and editing.
+
+Dated local commands used supported Node22.18, nice10, one worker, in the
+existing clean-installed partial probe (not full-tree acceptance):
+
+```sh
+node node_modules/vitest/vitest.mjs run --root packages/web \
+  src/lib/__tests__/edgeRouteGeometry.test.ts --maxWorkers=1 --no-file-parallelism
+node node_modules/typescript/bin/tsc --noEmit -p packages/web/tsconfig.json
+node node_modules/@playwright/test/cli.js test tests/e2e/graph/item-wire-consumers.spec.ts --list
+node --test --test-concurrency=1 tests/lib/runner/unified-cli.test.mjs
+```
+
+Results: 11 geometry cases passed in0.722s (9ms bodies); scoped web types passed
+in3.72s; two browser cases discovered, **not executed locally**; 32 CLI harness
+cases passed in4.56s without a browser. Scoped flat-config lint over both
+components/helper/test passed with0 errors and314 warnings. An initial lint
+invocation from the uninstalled worktree failed to resolve a legacy config;
+the corrected command ran from the probe's web package with
+`ESLINT_USE_FLAT_CONFIG=true`. That setup error is not a source lint result.
+
+The new blocking `item-wires` sequence is enrolled in PR/full profiles and CI
+after smoke, before longer suites, with one browser worker and zero retries.
+It exercises actual SVG paint/hitboxes/arrows, pointer dragging, wire/label
+editing and initially hidden label sizing, retains PNG/JSON, rejects writes,
+compares original graph data and uses owned fixture cleanup. CI execution and
+artifact review remain required. The route is still straight; these changes
+do not establish bounded obstacle routing, visible detail budgets, complete
+spaghetti detection or closure of #220.
+
+Independent review of `9931b3c` required stronger browser oracles before
+acceptance: pinned physics can be at rest while the450ms Fit transition is
+still moving the camera, and paint checks must inspect ancestors and actual
+stroke/fill. The correction waits for both `atRest` and `cameraReady`, rejects
+hidden ancestors/absent paint, and deliberately hides wire and arrow groups or
+removes a wire stroke to prove rejection. It also compares route endpoints to
+the original API-directed cards and checks the editor's original direction and
+selected type. Relationship type buttons expose their selected state through
+`aria-pressed`. Local discovery still finds2 cases and scoped web types pass;
+actual browser acceptance is pending. These were source-review findings,
+not locally executed browser negatives.
+
+On2026-10-28, exact source `93fe6d2` CI40646515194 failed the hidden-size case:
+expected at least104, received94.90625. The other item-wire case passed,
+including paint negatives and actual dragging/editing. Chromium had returned
+valid measured text dimensions smaller than the conservative estimate; the
+assertion incorrectly assumed hidden SVG always measures zero. This is a test
+precondition defect, not evidence that valid94.90625-unit geometry clips text.
+The amended case deliberately returns an empty box only for hidden edge text,
+counts that branch and retains the full-size/background/visibility assertions.
+
+The owner then requested a visible headed browser. A single installed Chrome,
+one worker, zero retries, on isolated preview28917/API28918 and the previously
+owned Neo4j28687 exercised the regression. With only the positive-measurement
+production guard removed from the disposable preview, the actual case failed
+with width36 versus104, after confirming an empty measurement. Restoring the
+unchanged production guard made both item-wire cases pass in21.9s. Traces,
+screenshots and logs are retained under `item-hidden-negative` and
+`item-wires-headed-positive` in the20261027 evidence directory. This reused a
+matching dependency installation for scoped preview evidence; it is not a
+clean-install/full-tree acceptance run. The source PR remains subject to fresh
+CI and independent review. The owner's further feedback about clipping,
+overlap, tiny text and wasteful cards requires broader visual acceptance;
+passing these two interaction cases does not resolve that feedback.

.agents/inbox/2026-10-28-offset-label-drag-needs-offset-path.md

@@ -0,0 +1,47 @@
+# Project dragged labels onto their displayed offset path
+
+2026-10-28 UTC, detailed item-wiring preparation for #220. Existing guidance
+was searched through `.agents/issues/INDEX.md`, `edgeLabelLayout.ts` and the
+schematic lessons. This is a controlled helper regression found before D3
+integration, not an observed failure in a deployed item router.
+
+The independent review of `eee4063c43e619a33c48fd0408bb9dc2dd45cc92` asked for
+an offset-label drag case beside a close return wire. The helper initially
+projected the pointer onto the wire centerline. On a U-shaped route whose
+horizontal legs are 40 units apart, the lower leg's upright label is offset22
+units upward. Its center is18 units from the upper wire and22 from its own.
+Nearest-wire projection therefore selects the wrong leg even with the prior
+fraction available as a tie-breaker: the distances are not tied.
+
+Command, supported Node22.18, existing clean-installed partial probe, one worker:
+
+```sh
+node node_modules/vitest/vitest.mjs run --root packages/web \
+  src/lib/__tests__/edgeRouteGeometry.test.ts --maxWorkers=1 --no-file-parallelism
+```
+
+Before correction: 9 passed, 1 failed, exit1, 0.713s. The executed new assertion
+expected `340 / 440 = 0.772727...` but received `100 / 440 = 0.227272...`.
+It passed a requested offset22 while the earlier helper ignored that extra
+argument. The production API was then extended explicitly, not the assertion
+relaxed. After correction: 10 passed, no skips, exit0, 0.691s (8ms test bodies).
+Scoped strict ES2020 TypeScript checking of helper and test also passed. Source
+and test were copied unchanged into the probe; this is not full-tree acceptance.
+
+Corrective rule: use the same upright perpendicular displacement for pointer
+projection and label painting. At D3 drag start, retain where inside the label
+the pointer grabbed it, then project the adjusted label center onto the offset
+route. Preserve icon rotation around the icon's own center. Keep the direct
+wire projection available for wire hit testing; it has different geometry.
+
+The review also found repeated route scans and candidate sorting in placement.
+Interpolate each candidate directly on its current segment and retain the best
+candidate in one pass. Placement now has O(segments × obstacles + segments)
+work with at most four candidates per segment. Runtime callers still need
+explicit bounds and a suitable update cadence. A null sampled placement means
+no tested slot was clear, not that every mathematical position is obstructed.
+
+The regression lives in `edgeRouteGeometry.test.ts`. No browser/database/image
+suite ran locally, no canonical positions changed, and no item-renderer
+acceptance is claimed. Carry this finding with the D3 consumer fix and its
+actual pointer/browser gates through normal review.