fix(agent): keep interruption feedback until it can be applied - #222

Describe what this PR does / why we need it

Contributing to ARGI: interruption feedback is no longer lost when InterruptionHook runs while the last assistant message still has pending tool calls.

Does this pull request fix one issue?

Fixes #211

Describe how you did it

• The hook reads the feedback without removing it, validates it, and only claims it with remove(key, value) when it is applied.
• With pending tool calls, valid feedback stays pending and is applied on the next model pass. Invalid feedback is discarded with a warning and never replayed.
• If newer feedback is written between read and claim, the conditional remove fails, the older value is not applied and the newer one is kept. The single-slot "latest value" semantics are unchanged.

Describe how to verify it

./mvnw -pl :argi-agent-framework -am test (1388 tests pass). New InterruptionHookTest cases cover deferral and single application, invalid feedback with pending tool calls, and a latch-controlled concurrent write; the deferral and concurrency cases fail on the previous implementation.

Special notes for reviews

Cleanup of ReactAgent.threadIdStateMap entries is out of scope and can be a follow-up. This only covers the in-process feedback path, not persistent approvals (#130).