Edit 1. Control Summary and Evidence Map Confirmed baseline: 124 icons for navigation, dashboards, alerts, exports, collaboration, and account administration; responsive web use at 16 px and 20 px; 24 px master grid; 2 px stroke; round caps/joins; 2 px corner radius; 2 px optical padding; fill reserved for selected states. Release goals are lower ambiguity and migration risk, preserved component APIs, and delivery within two sprints. Material risk Evidence status; blocked conclusion Smallest verification step Severity Duplicate concepts may trigger wrong actions or status interpretation. Missing: candidate list and usage captures; cannot decide merge, alias, deprecate, or keep separate. Supply candidate pairs with names and one context screenshot each. High Geometry may violate declared tokens and fail at 16 px. Missing: SVG viewBoxes, attributes, and path bounds; cannot size token drift. Export width/height, viewBox, path bounds, stroke, cap, and join for all assets. High Selected fill may overlap with alert, disabled, or destructive grammar. Missing: state pairs; cannot verify non-color differentiation. Supply default/selected/disabled/alert pairs at 16 px and 20 px. High Semantic fixes may conflict with API preservation. Confirmed constraint; missing mapping: public names, aliases, imports, and usage counts. Provide API inventory and consumer usage counts. High Two-sprint capacity may be insufficient if defects are systemic. Reasonable inference: 124 assets and six domains create broad reach; affected counts are unknown. Run inventory and geometry checks, then count defects by family and API reach. Medium Five evidence gaps: full name inventory; SVG measurements; product screenshots; duplicate-candidate list; component API and usage data. These gaps block icon-level scoring, accessibility findings, platform-conflict claims, and migration sizing. When an asset conflicts with a token, record both values. Controlling assumption: the declared token governs unless a documented product convention or 16/20 px legibility test justifies a bounded exception. 2. Taxonomy and Semantic Audit Consolidated finding Classification Evidence and conclusion Risk rank; path Navigation and wayfinding Role: destination, hierarchy, disclosure, panel control. Families: home, dashboard, folder, menu, arrows, chevrons. States: default, selected, expanded, collapsed, disabled. Scope confirmed; assignments missing. Filled selection is applicable, but destination icons, disclosure controls, and panel toggles cannot be checked for overlap. High harm and reach; reversible if APIs remain. Separate destination from local control; review arrow/chevron/caret candidates in context. Dashboard and data controls Role: visualization, filter, sort, drill, refresh, compare, inspect. Families: chart, table, funnel, sort, zoom, time, refresh. States: default, active, selected, unavailable. Reasonable inference: dense 16 px use makes recognition critical. Names and captures are missing, so chart variants, filter/funnel, and refresh/sync clarity are unverified. High harm; broad reach; medium confidence. Preserve distinct meanings for view, filter, sort, refresh, and drill; merge only when action and outcome match everywhere. Alerts and operational status Role: notification, warning, error, success, mute, snooze, acknowledge. Families: bell, triangle, circle-status, shield, clock. States: read/unread, active/muted, warning/error, acknowledged. Scope confirmed; state assets missing. Potential contradiction: declared "fill = selected" versus any filled alert/status asset; observed values are unavailable. Highest harm; reversibility varies. Do not encode both selection and severity by fill alone; verify silhouette, interior mark, label, and accessible name. Export and collaboration Role: export, download, upload, share, link, invite, comment, mention, assign. States: enabled, processing, complete, failed, restricted. Scope confirmed; candidates unconfirmed. Likely review areas are export/download, share/external, comment/message, invite/add-user; none may be called defective without evidence. Medium-high harm; mostly reversible. Decide by workflow outcome, not visual similarity; preserve separate API names where outcomes differ. Account and administration Role: settings, user, role, permission, organization, billing, security, delete, sign out. Families: gear/sliders, person/group, key/lock/shield, card/receipt, trash/exit. Scope confirmed; usage missing. Permission/security, organization/group, remove-user/delete-account, and sign-out/exit cannot be distinguished from supplied evidence. High harm; low reversibility for destructive/security meanings. Require labels and context-specific accessible names; do not merge by shape alone. 3. Visual, State, and Accessibility Scorecard All scores are Blocked, not failed: referenced assets were not included. Test area Declared control Blocked result and verification Severity Grid, bounds, optical padding 24 px grid; 2 px padding; 16/20 px output. No viewBoxes or path bounds. Compare bounds with 2 px inset, record intentional overshoot, and render pixel-aligned at both sizes. High Stroke, caps, joins, corners 2 px stroke; round caps/joins; 2 px radius. No SVG attributes or renders. Extract attributes, inspect expanded strokes, and compare effective corners after scaling from 24 px. High Negative space and legibility Dense dashboards at 16 px and 20 px. No small-size captures. Inspect closed counters, adjacent strokes, clipping, blur, and one-pixel collisions at 1× and 2×. High State differentiation Fill reserved for selected states. No state matrix. Remove color and compare selected, warning, error, disabled, blocked, muted, and processing states; verify focus and accessible naming. High Metaphor, convention, localization, duplicates Lower ambiguity; preserve APIs. No names, labels, locale contexts, guidance, or candidates. Review name + label + action + screenshot; test RTL directionality and unique accessible names. High If observed geometry differs from the token, show "declared 2 px" and "observed X px." The token controls acceptance unless the exception has an owner, rationale, bounded scope, and successful 16/20 px verification. Missing evidence never counts as a pass. 4. Prioritized Remediation and Migration Plan Action Family and required change Dependency Owner Verification Class 1. Build evidence ledger All families. Record role, family, states, risk, metrics, screenshots, duplicates, API name, aliases, and usage; no asset change. All five evidence gaps. Design-system lead + frontend engineer 124/124 complete records; unknowns marked missing, not inferred. Quick repair 2. Repair token drift without API change Any icon failing grid, bounds, stroke, cap/join, radius, or padding. Change SVG geometry only; preserve names/imports. Automated SVG audit and 16/20 px renders. Icon designer + frontend engineer Attribute lint, path-bound check, visual diff, small-size review. Quick repair 3. Separate selection from severity Navigation selected states and alert/status families. Keep fill for selected; use shape/interior-mark differences for warning, error, muted, blocked, disabled, and processing where applicable. State matrix and screenshots. Product designer + accessibility specialist + QA Grayscale pair review, focus check, accessible-name inspection, screenshot regression. Controlled migration 4. Consolidate confirmed duplicates behind aliases Dashboard, export/share, collaboration, and administration candidates whose action and outcome are identical. Use one canonical asset; retain old names as aliases. Context decision and usage counts. Design-system lead + platform owner Alias tests, bundle snapshot, no broken imports, no semantic loss. Controlled migration 5. Defer true semantic breaks Destructive, security, permission, or otherwise distinct concepts needing new metaphors or names. Do not force into two sprints unless critical harm is confirmed. Evidence, owner approval, consumer inventory, release plan. Product design director + engineering lead Task validation where available, sign-off, migration test, rollback plan. Breaking redesign Sprint order: Sprint 1 completes the ledger, automated checks, and low-risk geometry repairs. Sprint 2 completes high-confidence state fixes and API-stable aliases. Breaking redesigns remain deferred unless a critical operational, destructive, security, or permission ambiguity is confirmed. 5. Acceptance Tests and Stop Rule Inventory has exactly 124 public entries with canonical name, aliases, role, semantic family, states, and risk. Every SVG records declared and observed viewBox, bounds, stroke, cap, join, corner construction, and padding. Every token conflict records both values, controlling assumption, owner, rationale, and exception scope. Every icon renders at 16 px and 20 px without clipping, unintended blur, collapsed counters, or one-pixel collisions. Each used icon links to a product screenshot covering navigation, dashboards, alerts, exports, collaboration, and administration. Applicable default, hover, focus, selected, disabled, processing, warning, error, muted, and blocked states are tested; absent states are marked not applicable. Selected states follow the fill rule; no severity or availability state depends on the same fill treatment alone. Distinct actions have distinct labels or accessible names; shared geometry is allowed only when action and outcome are identical. Directional icons are reviewed for RTL; localization-sensitive metaphors are verified or explicitly blocked. Every duplicate candidate has a merge, alias, keep, or redesign decision with evidence, API impact, owner, and rollback path. Changed assets pass lint, visual regression, keyboard/focus review where interactive, and context QA at both target sizes. Release notes list repairs, aliases, deprecations, deferred redesigns, exceptions, and verification owners. Release-blocking evidence: the name inventory, SVG measurements, screenshots, duplicate list, and API/usage data are absent. This blocks icon-level classification, visual scoring, duplicate decisions, accessibility-semantic verification, migration sizing, and release approval. Stop rule: stop only when every material finding has severity, evidence status, affected family and count, remediation path, owner, and linked acceptance test. With the supplied evidence, the correct outcome is a controlled release hold: the audit method is complete, but no asset-level pass is supportable.