fix(interrupt/auth): prevent /stop swallow (incl. Bedrock) and empty-provider credential corruption

[kshitijk4poor]

Summary

Fixes two real bugs, salvaged from #60018 by @isheng-eqi (cherry-picked to preserve authorship), plus a follow-up that completes the interrupt bug class.

Scope note: the original PR branch #60018 was cut from an unrelated feature branch and bundled ~17 undisclosed commits (personality prompt #58774, copilot #59845, kanban, cron, file-tool validation, aux TTL, etc.) touching 18 files, while its body describes only 2. This PR salvages only the two in-scope commits that match the stated fix; the unrelated work belongs to separate PRs and is dropped.

Bug 1: /stop silently swallowed on streaming paths

Root cause: In the interruptible poll loops, the worker thread can exit before the main poll loop observes _interrupt_requested. The worker's stream callback breaks out and returns a partial response without raising, so the in-loop raise InterruptedError never fires and the loop returns the partial — /stop is silently swallowed.

Fix (original): post-worker _interrupt_requested re-check on the main OpenAI/Anthropic streaming loop.

Fix (follow-up, this PR): the Bedrock Converse streaming loop has the identical bug class — stream_converse_with_callbacks(on_interrupt_check=...) breaks and returns a partial SimpleNamespace without raising (bedrock_adapter.py). Added the same post-worker re-check there. The non-streaming loop (interruptible_api_call) is structurally immune and intentionally left unchanged: its only early worker return fires off _request_cancelled, which the main loop sets immediately before it raises in-loop, so no swallow window exists.

Bug 2: empty provider bypasses credential guard → corrupted agent state

Root cause: recover_with_credential_pool() guarded cross-provider pool swaps with if current_provider and pool_provider and current_provider != pool_provider. When agent.provider == "" (a valid unset state from agent_init.py:420 provider = provider_name or ""), current_provider is falsy, the guard is skipped, and the pool swaps base_url/api_key onto the agent without setting provider (_swap_credential never writes agent.provider), leaving provider="" model="".

Fix: if pool_provider and current_provider != pool_provider — an empty agent provider is now treated as a mismatch, so recovery is skipped (returns recovered=False, logs the mismatch) instead of corrupting the agent. This is the correct trade: recovery can't repair a missing provider, so skipping surfaces the error honestly rather than masking it. No sibling swap-site shares the vulnerable truthiness pattern; the custom custom:<name> normalization still holds when the provider is empty.

Tests

• tests/agent/test_bedrock_interrupt_post_worker.py (new): flips _interrupt_requested True mid-stream (after the pre-flight check) and asserts InterruptedError is raised on the Bedrock path; verified RED without the fix (DID NOT RAISE) and GREEN with it. Plus a sanity test that a non-interrupted call returns normally (guard doesn't fire spuriously).
• tests/agent/test_credential_pool_routing.py, tests/run_agent/test_credential_pool_interrupt.py: updated mocks to set pool.provider="" (MagicMock's truthy child would otherwise mask the new guard).

581 passed locally (excluding test_bedrock_runtime_default_region, a pre-existing environment-sensitive region assertion unrelated to this change — fails identically on clean main).

Supersedes #60018.

Co-authored-by: isheng ishengeqi@163.com