fix: keep failed turns in saved history from breaking the chat

[asasemahmed]

What changed

When one chat turn fails (for example, a provider rejects the API key), the main conversation stays broken afterwards:

• Reopening the app shows "Could not load conversation. Your saved messages have not been changed. <old error>", and the composer stays disabled on every visit.
• Past that, the stale error also appears as a banner with Retry response.

Connecting to a Rich Threads conversation replays its saved events, including the RUN_ERROR from the earlier failed turn. CopilotKit core reports replayed run errors through onError with code agent_run_error_event. hydrate() treated any error during connectAgent as a failed load, and the chat-wide onError subscriber showed it as a new error.

Now:

• runConversationTurn accepts error codes to ignore.
• While history replays, hydrate() and the chat error subscriber ignore only agent_run_error_event.
• Connection failures (agent_connect_failed) still block loading. Errors from new turns still stop the run and pause the follow-up queue, as before.
• Loads still replaying are counted, so when the thread changes mid-load, the old load finishing cannot end the new load's replay and bring back a stale banner.

Behavior change: if a live run fails while a reload is reconnecting to it, that error is now swallowed like a replayed one. No banner is shown; the failed turn stays in the transcript.

Verification

• Two new tests in tests/conversation-sdk.test.ts use the real CopilotKitCore:
  • A connect that replays a saved RUN_ERROR rejects without the ignore list, which reproduces the bug, and resolves with it.
  • A connect that fails still rejects.
• pnpm typecheck passes. biome ci passes on the changed files.
• Rebased onto main. pnpm test: 200 passed, 0 failed, 1 skipped (POSIX-only).
• Web app, before the fix: a thread whose previous turn had failed showed the "Could not load conversation" error on every load. After the fix, it loads with no stale banner, and new turns reply normally.

Integration limits

• If the page reloads while a run is in progress and that run then fails during reconnect, its error is not shown as a banner. The failed turn stays in the transcript.
• Tested on web only. The change is platform-independent state handling in chat.tsx.

[davidmckayv]

Needs a fix before merge. The replaying flag is one shared ref across effect runs. If the thread changes while history is loading, the old load's finally clears it while the new load is still replaying, and a stale error banner comes back. Tie the flag to the effect's active token or a counter instead. Also state in the description that a live run failing after a reload is now swallowed. #59 conflicts with this PR and should rebase after it.

[asasemahmed]

The browser-container failure is an unrelated flaky network test (it passed on #55 and #60); could you re-run it?

[asasemahmed]

Addressed the review: the replay flag is now a per-load counter, and the description states that a live-run failure after a reload is swallowed. CI is green now.

[jerelvelarde]

Value: historical failed turns no longer permanently prevent loading saved conversations. Filtering retains connection failures and new-turn failures. Three real SDK regressions passed. Template: clear verification/platform limits; please add the required UI screenshot or recording. The disclosed live reconnect-error banner limitation is understood. Updated-main diff reviewed; no actionable security finding. Merge after required CI.