fix(memory): forward on_pre_compress return value to compress() as pr…

[buptwz]

Problem

Fixes #23367.

MemoryProvider.on_pre_compress() lets external memory backends return a hint string telling the summarisation LLM which facts to preserve
during compression. The return value was silently discarded — every provider implementing this hook was being ignored.

Root Cause

conversation_compression.py called on_pre_compress() but threw away the return value. ContextCompressor.compress() had no
provider_context parameter, so the hint had nowhere to go even if captured.

Solution

• conversation_compression.py: capture on_pre_compress() return value, forward to compress()
• context_compressor.py: add provider_context param to compress() / _generate_summary(); inject as MEMORY PROVIDER CONTEXT block
  in the LLM summarisation prompt
• holographic plugin: add reference on_pre_compress() returning top-trust stored facts

Tests

10 new tests, all pass. Mutation-verified.

• tests/agent/test_compress_focus.py (4): provider_context injected into LLM prompt, empty string skipped, coexists with focus_topic,
  forwarded through compress()
• tests/agent/test_memory_provider.py (3): hint text passes through MemoryManager, empty and whitespace-only returns handled correctly
• tests/run_agent/test_compression_boundary_hook.py (3): end-to-end integration — hint flows from on_pre_compress() into
  compress(provider_context=...) via real `AIAgent

Tested on: Linux (Ubuntu 24.04, Python 3.13)

[alt-glitch]

Competing PRs for the same on_pre_compress() return value bug (#7192): #7195, #11236, #21118, #22576. This PR also references #23367. All address the same root cause — on_pre_compress() return silently discarded.

[teknium1]

Thanks for tracing the return value through the memory-provider and compression layers. The core premise remains valid on current main: agent/conversation_compression.py:634 still discards MemoryManager.on_pre_compress()'s result, despite agent/memory_provider.py:226-228 defining that return value as summary-prompt context.

Problems

• Current compaction now derives summary_focus_topic before invoking _generate_summary() at agent/context_compressor.py:2997-2999; salvage must preserve that behavior while adding provider context, including the retry calls at agent/context_compressor.py:2178 and 2195.
• agent/context_engine.py:87-105 does not advertise a provider-context argument. The existing TypeError fallback in agent/conversation_compression.py:640-644 reduces calls to current_tokens only, so a custom context engine that has not adopted the new argument will silently lose the hint and existing optional arguments.

Suggested changes

• Rebase the implementation concept onto the current compression call chain and cover auto-derived focus plus retry preservation.
• Make the context-engine compatibility path explicit and capability-aware rather than relying on the broad TypeError fallback.

Automated hermes-sweeper review.

[teknium1]

This fix has landed on main via PR #67938 (salvaged from PR #64342 by @GottZ, which consolidated the on_pre_compress return-value forwarding with sanitization, engine-signature compatibility, and lock hardening).

Your PR correctly identified the same root cause — the provider's return value being silently discarded before compression. The merged implementation supersedes this one, so closing it. First-submitter credit goes to @jairodriguez (#11236, Apr 16); thank you for the independent confirmation and fix.