fix: Return empty prerequisites for a flag that fails to evaluate in all_flags_state

[jsonbailey]

Summary

all_flags_state read result.prerequisites unconditionally even when a per-flag evaluation raised. The except branch set only detail, leaving result unbound (or holding a previous flag's value), so:

• the first flag raising → UnboundLocalError that aborts the entire state payload;
• a later flag raising → that flag silently inherits the previous flag's prerequisites.

Evaluator.evaluate() only catches its own EvaluationException; any other exception from malformed flag data propagates and reaches this path, so the except is genuinely reachable (rare, but real).

Fix

Bind prerequisites in both eval branches (success → result.prerequisites, error → []) so an evaluation error degrades only that flag. Adds a regression test (bad-first / good / bad-last) asserting no UnboundLocalError and that a failed flag does not inherit a neighbor's prerequisites.

This mirrors the same fix in the new async client (AsyncLDClient.all_flags_state).

Draft — surfaced during the async SDK review; no Jira ticket.

---

Note

Low Risk
Small, localized change to error handling in all_flags_state with a targeted regression test; no auth or data-path changes.

Overview
Fixes per-flag evaluation failures in all_flags_state so they no longer break the whole payload or leak another flag’s prerequisite metadata.

LDClient.all_flags_state now sets prerequisites in both the success path (result.prerequisites) and the error path ([]), instead of always reading result.prerequisites after the try/except. That avoids UnboundLocalError when the first flag raises and stops a later failed flag from inheriting a neighbor’s prerequisites.

Adds a regression test with bad-first / good / bad-last flag ordering to assert the call completes and failed flags omit prerequisites while a successful flag keeps its own.

^{Reviewed by Cursor Bugbot for commit 4f10055. Bugbot is set up for automated code reviews on this repo. Configure here.}