feat: DreamService 4th REM Vector — executeNLActionDigest()

Context

The DreamService REM pipeline (processUndigestedSessions()) currently executes 3 extraction vectors per undigested session:

1. Tri-Vector Extraction (executeTriVectorExtraction) — LLM-powered graph node/edge extraction from episodic memory
2. Topological Conflicts (extractTopology) — LLM-powered conflict detection (SUPERSEDES, OBSOLETES, DUPLICATE) → sandman_handoff.md
3. Capability Gap Inference (executeCapabilityGapInference) — Deterministic + LLM hybrid gap detection (DOC_GAP, TEST_GAP, GUIDE_GAP)

This ticket adds the 4th vector: executeNLActionDigest(), which reads successful Neural Link action sequences from the nl_action_log table (created by #9889) and uses them to close TEST_GAPs on graph nodes that have been verified through live agent interaction.

A2A Context (Fat Ticket)

Architectural Position

The 4th vector slots into processUndigestedSessions() after executeCapabilityGapInference() (line ~197) and before runGarbageCollection() (line ~211). It follows the exact same pattern as the existing three vectors:

1. Accept session context
2. Query a data source (nl_action_log via better-sqlite3 handle on memory-core.sqlite)
3. Perform deterministic analysis (group by sequence_id, calculate success rate per component class)
4. Upsert results into GraphService (TEST nodes with VALIDATES edges)

Key Design Decisions

• No Playwright scaffold synthesis in this ticket. The immediate value is closing TEST_GAPs automatically. Playwright synthesis is a follow-up concern.
• Success rate threshold. Only sequences where ≥80% of tool calls returned success: true should qualify as "verified interactions."
• Component class extraction. The nl_action_log.args JSON contains componentId and className fields from NL tool calls. These map directly to graph nodes of type CLASS.
• Gap closure mechanism. When a CLASS node has [TEST_GAP] in its capabilityGap property AND nl_action_log contains successful sequences targeting that class, the gap severity should be downgraded (not fully removed — agent interaction is weaker evidence than a permanent Playwright test).

Prerequisite

• #9889 (NL Action Recorder) must be merged first — the nl_action_log table must exist.

Files to Modify

Avoided Pitfalls

• The nl_action_log SQLite handle must use the same WAL-mode connection pattern as DatabaseService.mjs to avoid SQLITE_BUSY contention during concurrent swarm operations.
• Do NOT attempt to embed NL action sequences as vectors — they are structured relational data, not semantic text. Keep them in the SQL domain.

Verification Plan

1. Unit test following the DreamService.spec.mjs pattern (isolated tmp/ SQLite DB)
2. Mock nl_action_log with synthetic successful + failed sequences
3. Assert that TEST_GAP severity on corresponding graph nodes is downgraded after digest execution