The context engine was selected solely by the global `KernelConfig.context_engine`. An agent's manifest can now override it: a `[context_engine]` block in `agent.toml` (or `[agents.<name>.context_engine]` in a `HAND.toml`) is resolved as agent override > kernel config > compiled default, mirroring the existing per-agent `proactive_memory` / `skill_workshop` / `compaction` overrides (#5476).

`context_engine_for_agent` now picks the manifest override when present and falls back to the kernel-global engine otherwise (and on a per-agent build failure, so a malformed override never silently disables recall/compaction). Per-agent engines are built lazily and deduplicated by a fingerprint of their config, so many agents sharing one override config build a single engine. The plugin-allowlist gate now checks the resolved engine.

The new field is registered in `PER_AGENT_OVERRIDE_KEYS` (so a misplaced `[agents.x.context_engine]` in config.toml gets the #5476 "move to agent.toml" warning) and covered by the exhaustive override-key drift-guard test. No `KernelConfig` field changed, so config-reload classification is untouched; per-agent changes take effect on respawn, consistent with the other per-agent overrides.

Scoped to per-agent selection only. The proposal's separate `RequestLlmSummary` scriptable-hook contract changes the plugin protocol and is deferred pending design sign-off from the proposer.

…eltas

# Conflicts:
#	crates/librefang-kernel/src/kernel/tests.rs

…N, not a u64 hash

review follow-up: the override-engine cache was keyed by a u64 DefaultHasher digest of the config's canonical JSON, so a hash collision between two distinct configs would silently hand an agent the wrong cached engine. Key the map by the canonical JSON string itself — distinct configs can never collide, and the dedup (one engine per identical config) is unchanged. Negligible cost for tiny, few configs.

…oks (#6264)

Completes the second half of #6264. A scriptable compact hook can now return { request_llm_summary: { prompt?, summarize, keep, max_tokens? } } instead of final messages; the host runs the LLM summary with its own driver and assembles the CompactionResult (summary = LLM text, kept_messages = the script's keep list). The script chooses what to fold and how; credentials/routing/driver stay host-side. Handled at all three compact output sites (cache hit, cache miss, non-cached) via a shared try_host_summary helper. Missing/empty summarize or any LLM error falls back to the built-in compactor — never a panic. 3 tests cover the host-run path, the no-directive passthrough, and the empty-summarize fallback.