Instant: stop stale background sync when embedded views are disposed (workaround + stress example)

[mutumbakato]

Problem

Closing a NutrientInstantView on Android does not stop that document's listen-for-server-changes long-poll. The InstantPdfDocument survives in the descriptor cache with its sync coordinator still listening, so every document ever viewed keeps a background /sync connection open. After viewing enough distinct documents, these stale long-polls exhaust the per-host connection pool and the next document's auth/download queues behind them:

• Document loads stall for 40–60 seconds (measured 47.6 s stalls on a healthy local Document Engine, no fault injection).
• Behind a reverse proxy with a read timeout, the stale polls die as endless empty-body 502s for previously viewed documents.
• iOS is unaffected: it correctly stops sync on view teardown (verified side by side).

What this PR contains

1. InstantSyncLifecycleAdapter (in the new example) — a workaround using only public APIs and the stock pub.dev packages:

• At onPdfFragmentReady, disables listenToServerChangesWhenVisible before the document loads. Timing matters: an already-running long-poll holds its pool slot until the server releases it (~47–60 s measured), so stopping at dispose alone is not enough (we measured that variant: stalls unchanged).
• At onFragmentDetached, stops listening on the cached document (belt and braces).
• syncNow() pulls server changes on demand, replacing the disabled real-time listener. Local edits still upload automatically, and every push also applies server changes.

Registered once via platform.Nutrient.initialize(androidAdapter: ...). Verified against a licensed Document Engine that adapter registration does not affect licensing.

2. "Instant Switch Stress" example — automates document-switch / close-reopen cycles against a Document Engine and reports per-cycle load latency, with a UI toggle to compare baseline vs workaround.

3. Fixes example/pubspec.yaml — removes dependency_overrides pointing outside the repository (monorepo sync artifact that broke pub get on clean checkouts).

Measured results (12 distinct fresh documents per run)

Configuration	Median load	Stalls
Baseline	746 ms	47.6 s at cycles 6 and 11
Stop-at-dispose only	767 ms	unchanged (in-flight poll keeps its slot)
This adapter	~250 ms	none

Server-side confirmation: without the workaround, Document Engine logs show all previously viewed documents still re-polling /sync minutes after their views were disposed; with it, only the mounted document syncs.

Trade-off

While a document is passively viewed it no longer receives other users' changes in real time; changes arrive via syncNow(), on any local edit, or the periodic background sync. For single-user review workflows this is typically acceptable. The proper SDK fix (full sync cleanup on view teardown, matching iOS) is tracked internally and lands in an upcoming release.

Try it

cd example && flutter run
# open "Instant Switch Stress (repro #134164)", toggle the workaround, Run

The example defaults to a local Document Engine on :5001 with documents fresh-doc-01..30 and clean-doc-a uploaded (see the header of instant_switch_stress_example.dart for setup). To adopt the workaround in an app, copy InstantSyncLifecycleAdapter from the example and register it at startup — license initialization via Nutrient.initialize(...) is unaffected.