searchi/docs/EARLY-TEXT-SETUP.md

# Text readiness during media setup

The desktop can prepare the real Google query session as soon as the pinned
309,855,456-byte backbone is verified. The 554,821,024-byte media projector
continues downloading independently. Verification events follow exact SHA checks
and atomic installation; byte-count progress never substitutes for verification.
The complete model bundle still requires both artifacts.

The UI distinguishes warmed search, text/model readiness, media/model readiness
and indexing admission. A media-download error remains visible with a retry
action while the retained text query session stays usable. Worker capability
alone does not establish that the media projector is installed.

## Index coverage and promotion

Initial text indexing can run during projector download after query preparation.
Text-only status counts text passages, rather than repeatedly treating every
unindexed media file as a reason to rebuild. The automatic attempt key includes
requested coverage, so completed media setup can trigger full promotion.

A retained reader selects a compatible active publication inside the same SQL
read snapshot. Prepared-graph lookup uses that publication's actual fingerprint.
The query model and task-specific vector LRU remain loaded through promotion.
Model, dimensions and text-window identity remain strict.

Text tasks use a validated component policy shared by new text-only/full-media
generations. Cross-coverage donors are admitted only from the current active
generation after bounded metadata checks; the existing exact-policy donor
lookup across completed generations remains available. Media donors still
require the exact media generation policy. Completed chunks, exact extraction
bytes and persisted ranges must pass the existing provenance checks.
Historical text-only or custom-media text tasks using the previous
whole-generation policy can need a one-time re-encode; their lexical and
semantic readability is preserved.

Early setup never downgrades an existing compatible full media generation,
including a stale one during projector repair. The final decision is made under
the index build lock; a stale UI status snapshot cannot bypass it. Updates resume
after media setup is ready. Explicit CLI text-only builds retain their behavior.

## Actual validation

Windows VM tests force CPU. The genuine Google `early_text_google_cpu` regression
passed in **42.94 seconds** with the verified normal worker, two text sources and
one real image. It opened and warmed the reader before the first publication,
then verified early remembered-text retrieval and citations, prepared ANN handoff,
full image promotion with **zero document re-encodes**, identical text-vector
bytes and retained query-cache calls. A changed source was suppressed, and the
guarded text builder preserved the full generation. This is a correctness fixture,
not a cold-download, first-input, p95 or representative-library performance result.

Seven additional service control tests cover coverage, mutation, strict identity,
donor reuse and publication admission; they do not produce semantic-quality
evidence. Desktop state tests cover early text admission, media failure and full
index preservation. Frontend tests cover coverage promotion, retry visibility,
legacy status and incompatible-generation readiness.

The final Windows CPU workspace run passed **137 tests**, with **24 explicitly
ignored** integration/resource checks. Desktop Rust passed **32 tests**, frontend
passed **51**, and both strict all-target Clippy checks, formatting and the
frontend production build passed. The actual Google integration above was run
separately; ignored entries in the default suite are not execution evidence.

Native Linux UI interaction during a genuine delayed HTTPS media download,
failure, retry and full promotion passed on `9aae3f8`, retaining the query worker
and explicit selection. Its actual first unseen query reported 84 ms native and
293 ms through the DOM update on four fixture files. A later cached debug request
is a separate sample. The test exposed a semantic-activation clock defect;
the frontend now starts an independent refresh measurement, but a focused native
retest remains pending. See [native evidence](LINUX-DESKTOP-VALIDATION.md).
The owned loopback CONNECT proxy and isolated app environment do not change
the product's pinned download URL or verification. Native Windows/macOS, GPU,
large libraries and actual first-input measurements remain open. Document indexing
still needs to finish a publishable generation; earlier model readiness does not
make a large folder's content instantaneously available.