feat(memory): make external provider prefetch timeout configurable

[Navlem]

Problem

agent/memory_manager.py hardcodes _EXTERNAL_PREFETCH_TIMEOUT_S = 8.0, but the correct value depends on the external memory provider's recall latency — something the core cannot know.

This is not only a "memory arrives late" issue. In MemoryManager._prefetch_external (memory_manager.py:611-632) a timed-out worker thread is not cancelled:

thread.join(self._external_prefetch_timeout)
if thread.is_alive():
    logger.warning("Memory provider '%s' prefetch timed out after %.1fs; ...")
    return ""

The thread keeps running, and lines 611-619 skip prefetch on any later turn while it is still alive. So when a provider's recall reliably exceeds the timeout, prefetch is never applied at all — not merely delayed. The only symptom is a repeating warning, which reads like noise rather than "your memory provider is switched off."

Observed with the Hindsight provider against a bank whose local CPU cross-encoder reranked 300 candidates per recall: recall measured 15-22s against the 8s ceiling, 135 warnings accumulated, and prefetch was effectively disabled for the entire period. Raising the timeout restored it immediately.

Change

Adds memory.external_prefetch_timeout (default 8.0, so existing behaviour is unchanged) and threads it from agent_init into MemoryManager.

Deliberate choices:

• config.yaml, not an env var — this is a behavioural setting, not a credential.
• Invalid values fall back, they do not raise. A non-numeric or non-positive value in user config must not break startup, so it degrades to the existing default. MemoryManager's own positive-value guard is left intact for direct callers.

Tests

tests/agent/test_memory_prefetch_timeout_config.py:

• Resolution chain: configured float / int / numeric string, absent key, None, zero, negative, garbage, empty config.
• Asserts the documented default in config_defaults.py matches _EXTERNAL_PREFETCH_TIMEOUT_S, so the two cannot silently drift.
• Guards the real agent_init call site. The parametrised cases mirror agent_init's resolution logic, which means they would stay green if agent_init reverted to a bare MemoryManager() — so that case is asserted separately against the source.

Mutation-verified: breaking the wiring fails the guard test (exit 1); restoring it passes (exit 0). 88 tests pass alongside the existing tests/agent/test_memory_provider.py suite.

Verification

Exercised on a live install rather than only in tests: set to 15, restarted the gateway, confirmed the value resolves through load_config() → MemoryManager as 15.0, and confirmed real hindsight_recall calls now complete instead of being discarded. Zero timeout warnings from sessions created after the restart.

One behaviour worth noting for reviewers: the timeout is read when a session's agent is constructed, so already-running long-lived sessions keep their previous value until they end. That is consistent with how other memory.* settings behave.

[alt-glitch]

This was generated by AI during triage.

Duplicate of #87028. Both implement the same memory.external_prefetch_timeout config-to-MemoryManager handoff; #87028 is the earlier open and more complete implementation.

[rodrigogs]

I've posted the field evidence (Bedrock-only Docker gateway with a local Ollama embedder: 8s timeouts coinciding with 18.4-47.1s embeds and a 400 after 60s, 101/197 agent starts with no injected recall) on #87028, since the triage bot already flagged this PR as a duplicate of it and #87028 is the earlier, more complete head. Two of the same gaps apply here: cli-config.yaml.example's memory: block (ends at line 908) doesn't get the new external_prefetch_timeout key, and test_agent_init_actually_passes_the_timeout asserts on inspect.getsource(agent_init) text, which is brittle against the _init_memory refactor already on main (agent_init.py:1222-1290) - a call-based test (invoke _init_memory directly with a patched load_memory_provider and assert on agent._memory_manager._external_prefetch_timeout) would survive that. Not duplicating the full note here - see #87028 for the rest.

[Navlem]

Closing as a duplicate of open #87028, which covers the same external_prefetch_timeout registration/handoff with stronger validation. This is not fixed on main. The keeper still needs rebase, actual _init_memory wiring tests, sample/docs, and coordination with provider-specific defaults. Preserve configured/invalid/missing-value tests in that work.