fix(sse): honor disconnect/dispose and cap error-path reconnects

[cursor]

Bug and impact

Trigger: A screen subscribes to an SSE API (invokeAPI with listenForChanges / SSE provider). The user calls disconnectSSE, navigates away (screen dispose), or the endpoint errors repeatedly.

Impact:

1. Disconnect/dispose ignored: disconnect() and dispose() set _manuallyDisconnected but reconnect logic never checked it. dispose() also cleared the set immediately, so delayed onDone callbacks still scheduled reconnects against closed HTTP clients — background network activity, battery drain, and possible duplicate listeners.
2. Unbounded error reconnects: _handleSSEError received reconnectAttempts by value; increments inside the handler did not update the outer counter in _connectSSE. With autoReconnect: true (default), every stream error retried with attempt == 0, bypassing maxReconnectAttempts (default 5) and causing an infinite reconnect loop on persistent failures.

Root cause

• _manuallyDisconnected was write-only (set in disconnect/dispose, never read in onDone/_handleSSEError).
• dispose() cleared _manuallyDisconnected after cancelling subscriptions.
• Error-path reconnect used a by-value int instead of shared mutable attempt state.

Fix

• Add _disposed flag and central _shouldReconnect() guard (checks dispose, manual disconnect, autoReconnect, and maxReconnectAttempts).
• Re-check guard inside delayed reconnect callbacks before calling connect().
• Share reconnect attempt counter via List<int> so error and onDone paths increment the same counter.
• Stop clearing _manuallyDisconnected in dispose().

Validation

• Extended modules/ensemble/test/sse_provider_dispose_test.dart with reconnect guard tests (disconnect, dispose, max attempts, shared counter).
• Run from modules/ensemble: flutter test test/sse_provider_dispose_test.dart

Duplicate check

• Merged PR fix(sse): scope dispose to per-screen provider instance #2271 (cursor/critical-bug-remediation-d916): scoped SSE dispose to per-screen provider instances — related but does not fix reconnect-after-disconnect or error-path attempt bypass.
• DRAFT PR fix(action): restore page apiMap after executeAction completes #2275: executeAction apiMap restore — different code path.
• Open PR fix(firestore): dedupe listenForChanges subscriptions to prevent crash #2257: Firestore listen dedup — unrelated.
• Closed PR fix(cdn): retry manifest fetch after transient download failure #2262 / branch cursor/critical-bug-remediation-76ab: CDN _lastUpdatedAt premature commit — different bug, not re-opened here.
• No open PR, issue, or automation branch targets SSE reconnect guards or error-path maxReconnectAttempts bypass.