Prioritize displayed relationships with connected-pair fallback

.agents/inbox/2026-10-28-browser-evidence-needs-files.md

@@ -0,0 +1,51 @@
+# Browser evidence must exist as files before report collection
+
+2026-10-28, PR #224. Independent review comment4404454674 verified that
+`runPlaywrightSequence` ignores attachments without a path. The new priority
+journeys used `info.attach({body})`; the actual Playwright JSON normalizer retained
+body/no path, and the real adapter probe returned a passing case with zero
+attachments and zero errors. The517742c2,000-item stage completed successfully
+in CI40742884920 at18:22:57UTC, but this did not establish retained priority
+accounting evidence. Preserve that distinction and the earlier report.
+
+Cause: producer-side inline attachments did not meet the existing unified
+collector's file-path contract. A green browser assertion alone does not prove
+that another machine can inspect its measurements or screenshots.
+
+Corrective rule: write JSON via `info.outputPath` and `writeFile`, save screenshots
+to output paths, then attach by `path`. Required evidence must survive the actual
+unified report and uploaded archive, not only Playwright's in-memory report.
+Do not weaken owned-path checks or silently call missing evidence accepted.
+
+The corrected1440px headed priority journey passed in13.09s with one worker and
+zero retries. Its actual JSON report lists file attachments for
+`priority-accounting.json` (3193 bytes) and `priority-fallback.png` (61545 bytes),
+and both files exist. Retained local report: `priority-artifacts6.json` in the
+owned resolution evidence directory. The2,000-item case now writes and attaches
+`two-thousand-priority.json` by the same existing pattern; it is not rerun locally
+under the workstation resource constraint.
+
+Limits: this is a correction to the new test producers, not a claim that all
+legacy inline attachments are collected. Final exact-source CI must execute the
+large case and independently verify its retained counts, screenshots, artifact
+paths and source fingerprint. No runtime or policy behavior changed.
+
+Independent inspection of the materialized 9b8409c image found a second evidence
+gap: the screenshot showed an almost empty canvas after reload/inspection, even
+though canonical fallback IDs were correct. The test now waits for camera
+readiness, frames the owned pair, and requires its actual painted stroke and
+visible “Depends On +1” label. It uses the shared symbol oracle, including useful
+visible titles, minimum text/wire weight, and label/card/wire clearance.
+
+The first strengthened run (`priority-artifacts7`) rejected a painted horizontal
+SVG path because Playwright's area-based visibility predicate sees zero height.
+The replacement checks ancestor visibility, actual stroke/opacity/length and
+viewport intersection. The next run (`priority-artifacts8`) correctly rejected
+the 390px scene: its horizontal endpoint titles were both off-screen. The phone
+fixture now positions the same owned Gateway/Worker pair vertically before the
+API baseline is recorded. No assertion or source-position invariant was relaxed.
+Both desktop 1440 and phone 390 journeys passed in 25.17s in
+`priority-artifacts9`, with actual geometry JSON, PNG and accounting files. The
+phone image shows both endpoint titles and the fallback connection; the desktop
+image shows four cards. All failed reports/images remain retained. This remains
+a test/evidence correction; no product defect was established by these failures.

.agents/inbox/2026-10-28-edge-deletion-must-refresh-subscribed-counts.md

@@ -0,0 +1,62 @@
+# Refresh every subscribed edge query after deleting a relationship
+
+On 2026-10-28, inspecting the actual `priority-typography10` desktop/phone
+JSON exposed a count mismatch after the intentional BLOCKS deletion: canonical
+API data and the graph dock had four relationships, while the graph-selector
+header still read `4 nodes, 5 connections`. Existing cache/count lessons did
+not cover this query-variable mismatch.
+
+The stronger desktop assertion failed before the fix:
+
+```text
+priority-count11: 0 passed, 1 failed, 20.15 seconds, zero retries
+Expected: 4 nodes, 4 connections
+Received: 4 nodes, 5 connections
+Timeout: 5000ms (14 reads of the unchanged header)
+```
+
+The failed report, screenshot and trace remain in the owned evidence directory.
+The exercised command was the existing `relationship-priority.spec.ts` with
+`--grep 1440 --headed --workers=1 --retries=0`, using only owned graph fixtures.
+
+Cause: the selector observes `GET_EDGES` with a source-graph filter; the graph
+observes the same document with source-or-target filtering. The relationship editor had no mutation refetch configuration and its parent
+callback refreshed only the renderer’s variables. Apollo stores those root lists
+separately, so normalized entity updates alone did not remove the stale ID from
+the selector's list. The selector does not poll.
+
+The correction supplies `GET_EDGES` itself to the mutation's `refetchQueries`,
+on the editor’s explicit Delete action, so Apollo refreshes matching observed
+queries using their own variables. The
+installed `QueryManager.getObservableQueries` implementation was inspected to
+verify document matching. The editor awaits those refetches before reporting completion. Its separate
+flip workflow is unchanged; no graph membership, priority policy or persisted
+position changes.
+
+Corrective rule: when deleting records represented by independently filtered
+query lists, refresh the subscribed lists or explicitly update each affected
+cache field. Test the user-facing count as well as the canvas and server state.
+Do not add polling merely to hide a mutation invalidation gap.
+
+The desktop/phone priority journey now requires the header to change from five
+to four connections after the one authorized deletion, alongside original-ID
+fallback, rendered text/wire geometry, no extra mutations and unchanged nodes.
+Final committed-source CI and independent review remain required. This is a
+deletion-path correction, not proof of every other mutation or cross-client
+refresh path.
+
+An initial local attempt changed the graph canvas mutation instead of the
+editor’s actual Delete action. `priority-count12` still failed the desktop header
+assertion; that attempted source change was reverted. Its phone case separately
+failed an earlier keyboard-return assertion before deletion; that report and
+trace are retained for independent review. A passing subsequent count test alone
+does not explain or erase that separate failure.
+
+With the correction on the actual editor action, `priority-count13` passed both
+complete desktop/phone journeys in 25.22 seconds, zero retries. The retained
+final typography samples now show `4 nodes, 4 connections`; all original API,
+priority, focus and geometry assertions also passed.
+
+Independent review traced the separate count12 focus failure to stale label
+clearance after the relationship panel closes. Its deterministic failing-before
+regression and correction are in [the priority context lesson](2026-10-28-priority-controls-must-preserve-context.md).

.agents/inbox/2026-10-28-preview-styles-need-rendered-verification.md

@@ -0,0 +1,68 @@
+# Verify emitted typography after refreshing a long-running preview
+
+On 2026-10-28, the owned GraphDone headed preview still displayed some small
+text after the readability change. Searching the symptom index and existing
+readable-symbol, dependency-parser and preview lessons found no matching
+configuration-refresh incident.
+
+The preview and reviewed source had identical `packages/web/tailwind.config.js`
+bytes (SHA256 `ab33c1897cd34ea18ff04cd6ef644558b59254dcaeb6120225eef4951e8f3ff8`).
+Nevertheless, reading the actual browser CSS rules before restart returned:
+
+```text
+.text-xs { font-size: 0.75rem; line-height: 1rem; }
+.font-normal { font-weight: 400; }
+```
+
+The document root was 16px. Visible graph counts, connection status and sidebar
+labels consequently measured 12px. The long-running Vite process had received
+source updates; source-file identity did not establish emitted-style identity.
+Only the owned Vite process was stopped and restarted from its web package with
+the same loopback API proxy. Reloading the same headed browser then returned:
+
+```text
+.text-xs { font-size: 0.875rem; line-height: 1.25rem; }
+.font-medium { font-weight: 600; }
+.font-normal { font-weight: 500; }
+```
+
+No product source changed. This establishes stale preview styling, consistent
+with retained build/configuration state; it does not isolate a specific Vite or
+Tailwind cache implementation. A fresh independent exact-source PostCSS compile
+also emitted the intended rules in 1.37 seconds, using locked Tailwind 3.4.19,
+PostCSS 8.5.28 and autoprefixer 10.4.21. See [the issue evidence](https://github.com/GraphDone/GraphDone-Core/issues/222#issuecomment-6477709156)
+and [independent review](https://github.com/GraphDone/GraphDone-Core/pull/223#pullrequestreview-5688035698).
+
+Corrective rule: after copying build or typography configuration into an existing
+preview, restart that owned preview and read computed styles from real controls.
+Do not stop unrelated services, assume hot reload refreshed configuration, or
+equate matching source hashes with matching generated CSS. Keep one headed
+browser and bounded serial checks on the shared workstation.
+
+The existing desktop/phone relationship-priority journeys now measure the actual
+graph title, graph counts, visible Connections dock and Links control. They require the 14px/500 secondary
+baseline, 14px/600 heading and 16px/600 control, reject a deliberate 12px/400
+visible counter, restore its original style, and retain named before/negative/after
+computed samples in `priority-typography.json` using file-backed attachments.
+
+Limits: these samples do not audit every page or every font. Earlier local SVG
+geometry checks remain evidence for their explicit assertions, not for all app
+utilities. The preceding CI build passed but did not retain generated CSS bytes;
+neither source compilation nor screenshots are retroactive byte-level inspection
+of that CSS. The strengthened journey must pass on its final committed source.
+
+The first refreshed-style run, `priority-typography10`, passed both complete
+journeys in 25.28 seconds without retries. Actual JSON samples showed 14/600,
+14/500 and 16/600 before/after, and the injected counter was 12/400. Inspection
+also exposed a stale relationship count; its separate cause and correction are
+recorded in [the deletion lesson](2026-10-28-edge-deletion-must-refresh-subscribed-counts.md).
+
+The phone intentionally hides the header count with its existing responsive
+CSS. That sample verifies emitted styling and subscribed data, not visible
+phone-counter readability. The final test also requires the actual Connections
+dock to be visible and readable, records visibility for every named sample, and
+applies the 12px/400 negative to that visible dock on phones.
+
+Final focused run `priority-typography16` passed both full journeys in 25.81
+seconds with zero retries. Its JSON explicitly distinguishes the hidden phone
+header from the visible 14px/500 dock and records the visible negative target.

.agents/inbox/2026-10-28-priority-controls-must-preserve-context.md

@@ -0,0 +1,60 @@
+# Priority controls must preserve current focus and inspection
+
+2026-10-28, PR #224 / issue #220. Independent review of `5aaf6ec`
+reproduced two interaction failures with the actual DOM/handlers:
+- Open a “Blocks +2” pair, change Links to All and close the pair browser. The
+  connected SVG trigger loses its tabindex, so calling focus does nothing and
+  keyboard focus remains on BODY.
+- Click the new Links control on a graph with fewer than40 items. The document
+  outside-click handler clears the selected item and editing edge; the same
+  control inside semantic overview is protected by that existing boundary.
+
+Review evidence: PR comments4404185217 and4404202776, summary5687337487.
+The initial helper/component suite passed41 cases and routing passed10; those
+passes did not cover these transitions.
+
+Corrective rule: after attempting return focus, verify the actual activeElement
+and fall back to the visible Browse trigger if the original no longer accepts
+focus. Preserve unrelated focus. A presentation control owns its click boundary;
+changing detail must not also dismiss the current inspector/editor.
+
+`relationship-priority.spec.ts` now opens a pair, switches to All, keyboard-closes
+it and asserts Browse receives focus with an unchanged camera. It also inspects
+Gateway, opens the native select by pointer, changes All/Essential by keyboard,
+and checks the same inspector/camera remain. Both1440/390 headed journeys passed
+in25.0s with one worker and zero retries (`priority-focus5`). Original identity,
+API snapshot, explicit winner deletion/fallback and cleanup checks remain.
+A preceding local command used the preview as its own copy source, failed the
+copy, then executed the old tests; `priority-focus4` is not regression evidence.
+
+Limits: local probes reuse owned dependencies. Full exact-head quality/browser
+CI, actual2,000-item priority artifacts, independent correction review and final
+dev integration remain required. No claim of full item routing/bus completion.
+
+## Recompute label clearance when the browser panel changes
+
+The `priority-count12` phone journey later failed its initial pair-return focus
+assertion before any deletion. Independent trace inspection showed the pair
+label was still present by ID but had `display:none`: a delayed label-placement
+pass had treated the open relationship panel as an obstacle. Closing the panel
+did not invalidate placement, so the now-unobstructed label stayed hidden and
+focus correctly fell back to Browse. Review comment4404899816 records the trace
+evidence; accepting Browse in this case would hide the layout defect.
+
+A deterministic regression changes All then Essential while the phone panel is
+open, verifies the obstructed label is hidden, then presses Escape. On the old
+runtime `priority-obstacle14` failed after 16.23 seconds: the label remained hidden
+for all 14 reads over the 5-second assertion deadline. Camera/data assertions and
+expected focus remain unchanged.
+
+The label-placement layout effect now also depends on `itemRelationshipScope`.
+Mounting/unmounting/changing the panel recalculates its actual obstacle footprint
+before the existing next-frame focus restoration. This is bounded placement on
+a UI transition, not a new per-frame router. The All-level case still falls back
+to Browse when the original label intentionally loses its interactive role.
+
+The corrected `priority-obstacle15` run passed both full desktop/phone journeys
+in 26.03 seconds with zero retries, including the deterministic hidden-to-visible
+transition, original-label keyboard return, accurate post-delete counts and
+computed typography negatives. The final priority CI still must validate and
+retain these assertions on the published source.

.agents/inbox/2026-10-28-trace-fallback-needs-directional-scope.md

@@ -0,0 +1,68 @@
+# Trace fallback must respect direction and preserve useful screen space
+
+Date:2026-10-28. Core PR224, issue220. Existing guidance on control flow,
+readable symbols and priority context was checked first.
+
+## Evidence and cause
+
+Final source377f5a3 passed the small priority and2000-item/5000-edge stages, but
+CI40751802514 failed both nested semantic browser cases. Retained review:
+https://github.com/GraphDone/GraphDone-Core/pull/224#pullrequestreview-5688631798 .
+Desktop incoming Payments selected an original DEPENDS_ON bundle, then wiring
+intersected it with the global unordered-pair selection. The opposite-direction
+BLOCKS winner removed the only eligible incoming relationship. The phone's
+additional Links row reduced the measured overview capacity to one group.
+The process-evidence failure is not a third independent failing browser case;
+downstream required stages did not run.
+
+The new component regression failed on the unchanged runtime: expected
+`['fallback','reverse']`, received `['fallback']`;6 other cases passed.
+Command, using an owned dependency fixture:
+`vitest run src/components/SemanticGraphOverview.test.tsx --maxWorkers=1 --no-file-parallelism`.
+After correction, that suite plus relationshipPriority and semanticTrace passed
+19 cases in3.29s. A first command used a nonexistent vitest.config.ts and failed
+before tests; it is not behavioral negative evidence.
+
+A first responsive candidate passed both headed nested journeys in67.27s,
+zero retries/errors, but its actual390px screenshot showed Zoom to Fit covering
+part of the group-count button. That candidate is visually rejected despite
+the green tests. The correction retains the reserved right control lane,
+uses one line for the phone group/item counts, and reduces spacing without
+reducing the14/16/18px typography floors or44px touch targets.
+
+Final focused headed desktop/phone run passed2/2 in51.51s, zero retries,
+errors or skipped cases, with the unchanged readable-text negative and strengthened
+incoming-direction/control-clearance checks. Actual geometry JSON and PNGs show
+two cards and one painted incoming route at both sizes: no card intersections,
+unrelated crossings/overlaps, diagonal segments or small text. Partition at the
+incoming trace is183 total =1 drawn +132 internal +50 external deferred; the
+previously summarized incoming original is now represented in its explicit scope.
+Command: `playwright test tests/e2e/graph/nested-semantic-lod.spec.ts
+--project=GraphDone-Core/dev-neo4j/chromium --headed --workers=1 --retries=0
+--trace=retain-on-failure --reporter=json`, against an owned local API/database.
+The superseded67.27s candidate and full CI failures remain retained.
+
+## Corrective rule
+
+An explicit incoming/outgoing trace re-evaluates per-original-pair fallback
+among that direction's canonical external relationships, before routing or
+paging. Its additional eligible originals are represented while tracing;
+other global originals remain accounted for. Exiting restores the global
+priority partition. Do not switch the user's level to All, reverse originals,
+or use the strongest group-level bundle to suppress other member pairs.
+
+Maintain exact partitions for drawn, deferred, summarized, internal and pending
+records. A component regression covers the opposite blocker, incoming fallback,
+a lower-tier independent member pair, All, direction changes, exit and unchanged
+input. The existing actual browser regression continues to require two useful
+cards, a painted incoming dependency, readable text, clear routes and no writes.
+It additionally checks top controls do not cover one another. Inspect screenshots;
+a geometry oracle that checks only routes cannot establish control clearance.
+
+## Limits
+
+These changes concern group traces and controls, not arbitrary item obstacle
+routing or larger browser capacity. Smaller screens can still honestly defer
+content when finite space cannot fit two cards. A focused owned-fixture run is
+not a clean-install/full-CI acceptance substitute. Earlier failures remain in
+GitHub artifacts and final exact-source plus post-dev evidence is still required.