test(desktop): close channel validation latch race

[tlongwell-block]

Summary

• make the e2e-only channel-read seam claim one concrete get_channels invocation before its work starts
• hold only that invocation with a single resolver; later unrelated reads pass through undeferred
• reject re-arming while the identified validation read is held
• remove the test's unnecessary mid-hold re-arm, preserving strict hold/release and repair-order assertions

The identity hook remains inside the existing maybeInstallE2eTauriMocks() boundary and does not affect production bridge behavior.

Failure reproductions

At the base implementation, both adversarial orderings were induced before fixing:

1. steal before target: an unrelated get_channels between arm and invalidation consumed the one-shot; the test failed waiting for the intended validation hold.
2. join while held: after the validation read was held, re-arm plus an unrelated get_channels reproduced CI's release-cardinality failure exactly: received 2, expected 1.

Validation at a21db5455e47409d695783faa29e3a3b3c50338f

• both adversarial orderings now pass: the target claims the one-shot, and post-target traffic completes without joining its hold
• focused regression: 20/20
• full community-rail.spec.ts --project=smoke: 19/19
• desktop unit suite: 3,992/3,992
• pnpm check / Biome and file-size checks: passed
• pre-push hooks reran successfully: branch-skew, desktop-check, desktop-test (3,992/3,992)

Integration gate

This SHA is not authorized for integration until:

• Wren completes exact-SHA identity-model review, independent full-spec execution, and both adversarial injections; and
• this PR's own Desktop Smoke E2E (1) is green at this exact SHA.