# Current-head matrix decision — [#23479](https://git.cafe/pingdotgg/t3code/pulls/23479)

Inspected complete five-file committed diff in /tmp/23479-current-complete.diff before any new reviewer capture. Target head [377dd251be](https://git.cafe/pingdotgg/t3code/commit/377dd251be942831904f75c29a63d85c6b22d0cd). Recomputed merge-base against local origin/main: [3b95e7831b](https://git.cafe/pingdotgg/t3code/commit/3b95e7831b11b17627dff16843bc087bfca397ec). Reviewer checkout was still b065ad799c at initial inspection; no checkout or state mutation performed.

VERDICT: SUFFICIENT — MATRIX ONLY. R1–R4 ×1440/390 CSS px ×light/dark remains the required 16 primary cells, plus four paired R2 empty controls. No new product behavior or permutation follows from the intervening seeder/Knip changes. Preserve /tmp/23479-final-matrix.md as the detailed contract; this decision supersedes its head and complete-file-count provenance only.

Full diff: web ModelPickerContent initializer chooses Favorites only if a saved favorite passes the existing inclusion predicate; web ProviderModelsSection appends non-custom synthetic rows for absent saved slugs; committed fixture/seeder; Knip manual seeder entry. The intervening seeder changes decode comments/trailing commas through fromLenientJson and create userdata before atomic settings write. This requires the independent rebuild to execute the pinned current seeder, not another fixture matrix. Fixture values are unchanged.

R1: initial fallback, select/reopen/search/clear/keyboard. R2: genuine reported favorite, Favorites once, select/unstar/stale reopen, separate empty control. R3: genuine Big Pickle star then disable, zero reported models, missing unstar/reload/picker. R4: missing+reported favorite dedupe, no custom CRUD, missing unstar, real unstar/restar/reload. Filter N/A for eight genuine models. Baseline missing-row removal is BLOCKED, never called successful.

Every current after cell needs new exact-head captures and continuous recordings; source equality is not capture provenance. Reused corrected BASE media retain original base provenance. Full-frame RGB×8 and max-channel>16 comparisons, recorded viewport/theme/DPR/font/scale/scroll/data and incidental differences remain required. Actual decoded video PTS must supplement wallclock notes. Old R4 rebuild/video proof remains supporting old-head evidence and cannot be relabeled.

Independent current-head R4 desktop-light rebuild is deferred until capturer declares its server stopped. Use current seeder in isolated state, fresh native preview tab, measured CSS1440×900/light/DPR2/font16/scale1, matching scroll and neutral pointer20,80, dismiss update toast before capture; compare entire RGB image without mask/alignment and stop only own captured server.

No source edits, commits, branches, pushes, comments, PR edits, watches or PR polling. Web only. Current capturer server/state untouched.