VUH-2148 import evidence

The historical landing batch dd8ee0d0-fd33-463d-9241-e14cfe276a50 selected 363 modules. It reported 508.53 seconds summed import time, 623.97 seconds test wall time, and 947.34 seconds for the complete landing gate. Summed worker time is not wall time.

We reused those exact 363 paths with Vitest's import-duration instrumentation and collectTests: no tests or hooks execute. Forks, isolation, four workers and the serial remote project remain unchanged. All four collections completed with zero module or unhandled import errors, against the original 990a774a base plus each successive source patch. Host load varied; instrumentation also makes these figures incomparable with the historical uninstrumented gate.

The final collection was 21.0% lower in wall time and 22.0% lower in summed import time than baseline. This is an observed result, not a controlled causal estimate or a passing full gate. The protocol phase did not improve wall time over the preceding phase under changing load.

Causes and remaining cost

The baseline's slowest files were Discord bridge room recovery (10.204s), relay conversation questions (8.266s), conversation tail sharing (8.104s), evidence fetch (8.095s), and lead roster/device (7.930s). Broad service and settings graphs repeatedly load dependencies in each isolated file.

• The models.dev JSON contributed 124.78 seconds across 266 files at import. It now loads and validates on first catalog use, retaining synchronous API and memoization. It is absent from the later import-only graphs.
• Pi AI's root contributed 69.99 seconds across 265 files. After using narrow model utilities and removing the voice provider barrel, its final measured self time was 5.73 seconds across 235 files.
• Protocol's root contributed 45.16 seconds self time across 357 files. Narrow settings imports and direct sibling imports remove broad internal cycles; final self time was 27.36 seconds across 355 files. Most service tests still import the public root directly. Its browser bundle remains Node-free.
• Pi coding-agent remains in 293 files: 210.37 seconds self time at baseline, 250.40 seconds in the final run. Lazy image processing removes that eager edge, but other captain/session/tool imports still load the SDK. Do not attribute its remaining cost to the removed image edge or claim it vanished.

External self time includes that external module's native import subtree. Inclusive timings overlap and must not be added. These measurements show why per-file isolation multiplies an eager graph; they do not measure a safe isolation alternative. No isolation, assertion or timeout was changed.

Verification

Focused catalog, image, provider and voice contracts: 98 tests passed across four files (4.38s). All six affected packages passed typecheck. The first provider typecheck found a missing type import after moving the family helper; the correction and successful recheck are retained. Explicit touched-file formatting and lint passed. Real esbuild checks proved the protocol index bundles for the browser and the bundled lazy catalog loads and memoizes. Settings and protocol route contracts also passed 61 tests in four files (1.03s).

The import commits landed as 303d4f2f4, f298b0115 and 6033b47f9. The live priority capture after deployment records its separate exact-tree gate timing and admission proof. The first import request stopped at dead-code checking: Knip did not infer the required CJS declaration edge. The correction explicitly registers src/bundled-snapshot.d.cts, following existing declaration entries. A full dead-code check passes; runtime behavior and schema validation are unchanged. Raw collection diagnostics, the exact selection, source patch and check logs are linked by the evidence manifest and attached to VUH-2148. The baseline manifest retains the original profiling and historical gate evidence.