…ing and session persistence

- Wire _call_summary_llm to auxiliary_client with structured prompts
  (Goal/Progress/Decisions/Files/Next Steps) and L1→L2→L3 escalation
- Fix multimodal token counting in TokenEstimator (was str()-ifying
  image URLs into token count)
- Add max_store_size config + ImmutableStore.prune() with GC for
  unreferenced entries, auto-prune on ingest
- Add lcm_metadata field to gateway SessionEntry for cross-session
  DAG persistence (backward-compatible)

…ence, store pruning, DAM e2e

- 19 tests for _call_summary_llm wiring and L1→L2→L3 escalation
- 17 tests for tiktoken integration and multimodal token counting
- 14 tests for gateway SessionEntry lcm_metadata round-trip
- 20 tests for store pruning with max_store_size cap and GC
- 18 tests for DAM end-to-end pipeline (ingest→index→search→recall→compose→persist→reload)

…M, wire async compaction

- Replace global _engine_ref with contextvars.ContextVar for thread-safe
  concurrent agent isolation
- Split 714-line engine.py into focused modules: query.py (LcmQueryMixin),
  format.py (LcmFormatMixin), session.py (LcmSessionMixin)
- Bundle DAM plugin from ~/.hermes/plugins/hermes-dam/ into agent/lcm/dam/
  with proper absolute imports; plugin delegates to bundled code
- Wire CompactionAction.ASYNC with threading.Lock + daemon thread pool;
  async_compact() runs compaction off main thread with callback support

…async compaction

- 6 tests for ContextVar thread/asyncio isolation
- 10 tests for bundled DAM imports and functionality
- 34 tests for engine mixin split with backward compat
- 17 tests for async compaction thread safety, callbacks, and locking

…_warned

- Only add LCM tool schemas to tool surface when tools are enabled
  (fixes test_no_tools_never_injects CI failure)
- Initialize _context_pressure_warned in __init__ to prevent
  AttributeError before first compression

…ssion rebuild

- Auto-initialize DAMRetriever in LcmEngine when numpy available
- Auto-sync DAM with store every 5 ingests (configurable interval)
- Add sync_from_store() to DAMRetriever to break circular get_engine() dep
- Search uses DAM ranking when retriever has patterns, keyword fallback otherwise
- Track compaction metrics: total, tokens saved, compression ratio, level distribution
- format_metrics() for human-readable stats
- Persist DAM state + metrics in session metadata for cross-session continuity
- Harden rebuild_from_session: validate pinned IDs, handle missing/null fields,
  degrade gracefully on corrupt data

…ild hardening

- 18 tests for DAM auto-init, periodic sync, search integration, session persistence
- 17 tests for compaction metrics tracking, format_metrics, async coverage
- 10 tests for rebuild edge cases: invalid pinned IDs, missing keys, corrupt data

…test

reset_session_state() was missing _user_turn_count and
context_compressor._previous_summary resets, failing 4 tests in
test_session_reset_fix. The toolset derivation test now patches
LCM_TOOL_SCHEMAS to keep its controlled tool set precise.

Cherry-pick holographic.py (phase-encoded HRR math), store.py
(SQLite fact store with entity resolution and trust scoring), and
retrieval.py (FTS5 + Jaccard + HRR hybrid search) from the
holographic-memory-store branch into agent/lcm/hrr/ following the
DAM bundling pattern. 12 tests verify imports, CRUD, and retrieval.

Compaction crystallizer: engine.compact() now auto-stores summaries
as persistent facts in the HRR store via _crystallize_to_hrr().
Failures are swallowed to never block compaction. 14 tests.

Session primer: prime_from_hrr() queries the HRR store for relevant
prior knowledge and injects it into the LCM active context at
session start. 13 tests.

Replace 11 fragmented tools (7 lcm_* + 4 dam_*) with 6 unified
memory_* tools that auto-route across all three memory layers:
- memory_search: DAM → HRR → keyword cascade
- memory_pin: LCM pin + auto-crystallize to HRR
- memory_expand: in-session or cross-session recall
- memory_forget: LCM forget + optional HRR trust reduction
- memory_reason: HRR compositional queries (probe/related/reason/contradict)
- memory_budget: token breakdown (delegates to lcm_budget)

Old lcm_* and dam_* tools remain functional for backward compat.
84 tests including schema validation, handler behavior, and compat.

…y_budget

- Warn when numpy unavailable and HRR weights redistributed
- Promote cascade search failures from debug to warning level
- Extract _normalize_ids() helper, removing duplicated normalization
- Validate and sanitize entities in memory_reason (strip, reject non-strings)
- Remove memory_budget passthrough (lcm_budget already covers it)
- Patch MEMORY_TOOL_SCHEMAS in toolset derivation test
- Wire HRR store init + memory_* tool dispatch in run_agent.py

The rebase onto v0.7.0 removed the ContextCompressor initialization
but v0.7.0 code paths still reference self.context_compressor for
token counting, fallback handling, and session state. Restore it as
a lightweight shim — LCM engine drives actual compression.