Command-picker test pruning — VUH-1925

Four mixed TUI files shrink from 2,332 to 2,123 test lines (−209). Five incidental cases are removed; the other 80 original cases remain, with presentation-only assertions pruned. No E2E, integration or golden test file is changed or removed. No timeout, fixture budget or selector is changed.

The ADR 0221 review ledger records each file's cut and keep reasons. Keep command/alias resolution, the Kitty key-release regression, keyboard selection/cancel/back/close, terminal width, multiline owner input and secret masking. Initial selection is checked by submitting it, rather than pinning a rendered marker or label.

Provider tests retain credential redaction and storage, hosted approval refusal, actual persisted provider/model/effort values, unreachable endpoint fallback, restart/provider intent and another face's later authoritative settings write. The shell keeps failed and uncertain prompts without losing newer drafts, steer/queue routing, streamed words, selected conversation IDs, close fallback, fresh context, side-conversation restoration and server interrupt handling.

Cut preview palette/spacer/copy, outline glyphs, status-row alignment, fixed auth-status tables and default slot labels, model presentation order, setup hint copy, conversation label/copy arrays and a constructor/default-state snapshot. The generic shell's duplicate ten-line tool-collapse assertions are removed; its user, assistant, reasoning and notice content assertions remain. Real native collapse/expand goldens remain in native-tool-render-integration.test.ts, and actual startup frames remain in fresh-page.test.ts.

Product code

Export-only, no behaviour change: newestFirst becomes private after its sole external snapshot consumer is removed. Its internal runtime caller and entire implementation body are unchanged. The proof compares the whole source file after removing exactly that one export modifier. Consumer searches also covered the private app and ops source trees. readinessFooter keeps its export because the real TUI entrypoint imports it. Knip stays active.

Root gate measurements

Both runs use the root gate's own native --changed selection. They return exit 0 and sourceStable: true. Neither is a zero-test clean-main pass.

These are not matched speedup measurements. The baseline checked HEAD 536263c92 against newly fetched base 289bd0d9b; that upstream change broadened its compiler scope. The after run is rebased onto 289bd0d9b; the export cleanup adds the existing setup-command and guided-setup integration files to native Vitest selection. Both guided-setup integration cases pass. Changing compiler scope, selected tests and shared-machine load prevents attributing the wall-time difference to pruning. Admission waits are excluded from gate wall time.

Measurements include exact HEAD/base/source fingerprints, phase times, file selections, per-package/repository source LOC and sampled load. Raw gate/Vitest/load receipts remain in the worker's ignored .local/vuh-1925/picker-{before,after}-* files. One baseline admission failed before child startup with the already-reported native lock OSError (errno 9); it is not a test or gate result. A later normal request completed successfully.

Retained higher-tier verification also passed unchanged: nine native rendering integration/golden cases and three actual startup-frame cases, 12/12 across two files in 1.21s of Vitest wall time. Receipt: .local/vuh-1925/picker-retained-vitest.json. The raw native JSONL and captured collapse/expand goldens were neither changed nor replaced with generated values.

The committed-head root gate, checked SHA, fixed base, retained native-golden verification and push receipt are attached to VUH-1925 after landing. The inventory remains incomplete; this batch does not close it.