fix(backend): stop mock modem re-broadcast clobbering merged client state (fixes flaky shard-2 E2E)

[andrescera]

What

In mock mode, broadcastFromDiff (apps/backend/src/modules/modems/modem-update-loop.ts) re-broadcast the full modem descriptor for every modem on every reconcile diff (broadcastModems(undefined)). This change makes mock mode use the same path as a real device: a full descriptor only for newly-added modems, a status-only partial for every changed modem. One file, +13/-9.

Why

The desktop shard-2 E2E job (modem-config-surface.spec.ts:100 — "a modem scan holds pending across same-list re-broadcasts…") failed byte-identically on every PR:

expect(locator).toBeDisabled() failed
Locator: getByRole('dialog').getByTestId('modem-scan-button')
Expected: disabled   Received: enabled

Root cause (a timing race, not a cross-file leak):

• The spec injects available_networks over the page socket via dev.emit, which is page-scoped and never persisted in the backend modemsState — so the backend's real available_networks for that modem stays empty ({}).
• The mock feedback tick (runMockFeedbackBroadcast, ~5s) re-sent the full modem state on every signal fluctuation, and buildModemMessage always includes available_networks (and config) in a full descriptor.
• The frontend merges each status.modems frame field-by-field (subscriptions.svelte.ts mergeModemList, shallow), so an incoming available_networks:{} overwrote the injected set.
• That flipped the scan signature and false-confirmed the pending scan, re-enabling the button.

The race only lands inside the 5s assertion window on slower/contended CI runners (backend up for minutes across ~25 prior shard-2 specs), so it passed locally but failed deterministically in CI. sim_lock is only serialized when present, which is why the sibling sim-unlock-surface.spec.ts survives the same tick; available_networks/config are always serialized, so they get clobbered.

A real device already sends status-only for a changed modem (the !shouldUseMocks() branch), so it never clobbers the scan list. The mock's full re-send was a stale artifact from before the frontend merged per-field. This aligns the two; real-device behavior is unchanged.

How to verify

• Repro: temporarily set MOCK_FEEDBACK_BROADCAST_INTERVAL_MS=250 and run modem-config-surface.spec.ts — it fails at the toBeDisabled assertion (byte-identical to CI). With this fix, -g "holds pending" --repeat-each=8 → 8/8 pass.
• Backend bun test: mock-feedback-broadcast, modem-available-networks-contract, modems-state-cache, modems-configure-applied, mock-modems-output, mock-monitor, monitor-manager, modem-migration, sim-state suites → 75 pass, 0 fail.
• E2E --project=desktop --workers=4 (grep-invert @visual|@a11y|@gallery): full suite → 226 pass / 3 skipped / 0 fail; modem/network/sim blast-radius subset → 25 pass.
• bun tsc --noEmit exit 0; biome check clean.

Risks

• Low. The change only affects the mock-gated broadcast shape; the real-device path is untouched. The initial full snapshot still reaches every client via the post-login buildModemsMessage() push and the discovery diff's added entries. The mock-feedback unit test asserts only .status.signal/.status.connection + dedupe counts, all preserved by a status-only partial.