# JEV-001 — Incorrect decision anchors and overlapping markers Status: Verification passed; preparing PR. Priority: highest. PR: Pending. ## Problem and evidence The monitoring reader initially highlighted its Sentry paragraph correctly. After later edits, both monitoring and Slack-alert markers highlighted the three Browser test checklist items. At 893×850 both marker rectangles were x=68.5, y=147.164, width=20, height=20. Clicking monitoring opened the overlaid Slack reader. Keyboard Enter opened monitoring correctly but highlighted the checklist. Record identity survived; passage targeting and placement did not. Observed in the [live browser pass](../jev-live-browser-test.md). Screenshot: `/Users/maggieappleton/.codex/visualizations/2026/10/02/01a0fd90-97a7-71b3-a895-3b2369b2570a/jev-live-anchor-regression.png`. The first corrupting mutation is unconfirmed; the sequence below is evidence, not a claim that every step is necessary. ## Concrete reproduction 1. Create independent monitoring and critical-error alert decisions; save Sentry and Slack. Verify each reader highlights its respective generated prose. 2. Reopen monitoring; manually add/select OpenTelemetry with Grafana and save. 3. Ask `@chopin` to append a three-item Browser test checklist. 4. Reload; create a temporary naming decision, then discard it. 5. Scroll at 893×850; inspect both markers, click and keyboard-activate each. The existing local document is `MaggieAppleton/elves/silver-storm` at port 8787. The original database/hosting decision must remain unchanged. Prefer isolated fixtures for destructive diagnosis. ## Expected behavior and scope Saved decisions keep their own prose relationships through choice replacement, unrelated Planner edits, appended/discarded cards, scrolling and reload. Each marker is accessible independently and opens its own reader. Missing or ambiguous passages orphan conservatively rather than highlighting unrelated blocks. Own the anchor lifecycle and only any proven marker interaction defect. Preserve main's shared reader/design system; do not restore the legacy Jev renderer. Exclude rationale generation and classifier/recovery policy (JEV-002/003). ## Acceptance checks - Reproduce the earliest corrupting mutation deterministically before fixing it. - Two saved decisions retain distinct correct highlights after each repro step. - Replacement rewrites the monitoring passage and preserves the Slack passage. - Unrelated Planner edits, append/discard, reload and scroll preserve associations. - Pointer and keyboard activation open the intended records; marker rectangles remain independently reachable at 893×850 and narrow/desktop layouts. - Missing/ambiguous anchors never silently attach to unrelated prose. - Unit/domain regression plus real-browser interaction and geometry coverage. - Independent tester and reviewer approve after any repairs; format/check results and a published PR image are recorded before handoff. ## Implementation and verification The isolated defect is target precedence: `resolvedKeys()` and the questionnaire store prefer question-level relationships even when a saved decision has its own valid prose anchor. A disposable fixture with two checklist question links and two distinct prose anchors reproduces both markers at the same y coordinate at 893×850. The focused unit regression likewise returns `checklist` instead of `monitoring-prose` before the fix. The shared editor now uses saved decision prose for resolved marker targets, highlights, jumps and related-block counts. Missing/orphaned saved prose remains inert instead of falling back to unrelated question context. Legacy records without saved prose keep their existing question relationships. No server, protocol, rationale generation or renderer replacement is included. Server regressions preserve separate prose anchors through two staged saves, monitoring replacement, unrelated Planner list append and persisted reload. The simple server sequence did not corrupt those anchors. Independent tester (Sol/medium) passed 22 targeted unit/domain tests, all four anchor-stress Chromium tests and three related prose lifecycle Chromium tests. The strengthened two-decision regression was then rerun after adding creation and discard of a third card through fake Jev chat and the shared Discard confirmation. It verifies distinct targets, exact highlight exclusions and pointer/Enter access before and after reload at 893×850, 390×844 and 1440×900. Related browser tests cover redecision replacement, prose deletion/orphaning and discard. Types, CI and diff checks passed; CI retains four existing lint warnings. Implementer (Sol/high) reproduced red unit target selection and red Chromium marker separation before the fix. Independent reviewer (Sol/high) found no production blocker in supported single-question Jev decisions and approved the settled production/test diff after the combined browser check. A narrow repair to the existing prose test helper tolerates the heading's existing “— show in plan” accessible-name suffix. Actual rendered UI image: `jev-001-correct-anchors.png`, captured after the combined temporary-card sequence; publication in the fix PR is pending. ## Remaining limits The exact mutation in the original live session is unconfirmed. Automatic approval review rejected a read-only query of its persisted sidecar/source as sensitive content extraction, so diagnosis uses disposable fixtures. The original prototype, database/hosting decision and existing checkout were not modified. No claim is made that every step of the paid-provider live sequence was replayed.