fix(plugin): preserve structure on PreCompact + capture compact summary

Two changes to the compaction flow:

PreCompact: agent stores its summary with infer=False
  on_pre_compact.sh now instructs the agent to call add_memory with
  infer=False. The agent has full conversation context and writes a
  structured summary; without infer=False the platform runs a second
  LLM extraction pass over that summary, losing structure and
  producing fragmented facts. With it, the structured text is
  preserved verbatim. Same guidance added to mem0-mcp/SKILL.md so
  the agent applies it to other already-extracted writes (decisions,
  anti-patterns, conventions).

SessionStart-compact: capture the platform-generated summary
  PreCompact fires BEFORE Claude Code generates its compact summary,
  so the summary is unreachable from that hook. SessionStart-compact
  fires AFTER, so a new background helper (capture_compact_summary.py)
  reads the transcript at that point, finds the isCompactSummary=true
  entry, and stores its content as a separate memory tagged
  metadata.type=compact_summary. The post-compaction bootstrap text
  now tells the agent to layer its recall queries across
  session_state (its own pre-compact summary), compact_summary (the
  platform's condensed view), and topical types like decision /
  anti_pattern.

Net effect: three complementary memory types from one session
boundary, none of which used to be filterable or even captured
correctly.

mem0-plugin/scripts/on_pre_compact.sh

@@ -5,9 +5,9 @@
 # the full context before it gets compressed.
 #
 # Output: Text instructions injected into Claude's context.
-# Claude still has the full conversation and can write an accurate summary.
-# A companion Python script (on_pre_compact.py) also runs to capture
-# transcript state directly via the Mem0 REST API as a safety net.
+# Claude still has the full conversation and can write an accurate summary,
+# which it stores via add_memory(infer=False) so the platform preserves
+# the structure verbatim instead of running a second extraction pass.
 
 set -euo pipefail
 
@@ -22,7 +22,9 @@ Context compaction is about to happen. You are about to lose most of your conver
 
 ### Step 1: Store session summary
 
-Call `add_memory` with a thorough summary covering ALL of the following:
+Call `add_memory` with `infer=False` and a thorough summary covering ALL of the following.
+
+`infer=False` is critical here: you've already done the extraction work yourself using full context. Without it, the platform runs a second LLM pass that loses your structure and pulls fragmented facts. With it, your summary is preserved verbatim.
 
 ```
 ## Session Summary (Pre-Compaction)
@@ -48,11 +50,19 @@ Call `add_memory` with a thorough summary covering ALL of the following:
 the post-compaction agent continue without asking redundant questions]
 ```
 
-Include metadata: `{"type": "session_state", "source": "pre-compaction"}`
+Tool call shape:
+```
+add_memory(
+  messages=[{"role":"user","content":"<the summary above>"}],
+  user_id="<the active user_id from the SessionStart bootstrap>",
+  metadata={"type":"session_state","source":"pre-compaction"},
+  infer=False,
+)
+```
 
 ### Step 2: Store any unstored learnings
 
-If there are learnings from this session that you haven't stored yet, store them as separate memories:
+If there are learnings from this session that you haven't stored yet, store them as separate memories with `infer=False` (same reasoning -- you've already extracted the fact, don't re-extract):
 - Failed approaches -> metadata `{"type": "anti_pattern"}`
 - Successful strategies -> metadata `{"type": "task_learning"}`
 - Architecture decisions -> metadata `{"type": "decision"}`