context-synthesize: bridge supplies wrong MemoryPattern fields and fabricates reward=1 (3.41.2) #3314

Narrow defect: context adapter fabricates outcomes and drops required fields

Reproduced with @claude-flow/[email protected], @claude-flow/[email protected], [email protected], Node 24.14.1. The real ContextSynthesizer works with its documented input; Ruflo supplies a different shape.

bridgeContextSynthesize() recalls hierarchical rows, then maps every row to:

{ content: r.value || r.content || '', key: r.key || r.id || '',
  reward: 1, verdict: 'success' }

AgentDB's ContextSynthesizer.synthesize() expects MemoryPattern[]: task: string, reward: number, success: boolean, optional critique, input, and output. It calculates success rate from success and extracts repeated patterns from critique; verdict and content do not supply these fields.

Controlled pure-function reproduction

Two synthetic episodes with the same meaningful critique, one success: true, reward: 0.8, the other success: false, reward: 0.2:

This test touches no database. The controller's default minimum pattern frequency is 2: one note producing no repeated patterns is not by itself a defect.

The important semantic boundary

Current hierarchical storage also accepts ordinary fact strings with no task outcome at all. Merely renaming verdict to success would fabricate successful episodes from facts. Empty/factual recall must not be misreported as measured success or quality.

Proposed minimal correction / acceptance

• Map only explicitly outcome-bearing, validated episode data to the documented contract, retaining both true and false success and the actual numeric reward.
• Preserve task and critique; validate optional fields before calling the controller.
• Plain facts and invalid/missing outcomes must be explicitly excluded/reported, or return an unsupported-input result when no eligible episodes exist. Do not invent reward or success defaults.
• Clearly distinguish no recalled rows, recalled but ineligible facts, partial eligible input, and controller failure.
• Test both hierarchical recall interfaces, asynchronous recall, mixed success/failure, zero reward, repeated critiques, plain prose, malformed JSON, inherited properties, and wrong-typed outcome fields.
• Do not rewrite retained memories, create a new memory store, or silently change hierarchy/consolidation behavior.

Related, but distinct

• AgentDB: 4 controllers unimplemented (ContextSynthesizer, ReasoningBank, BatchOperations, CausalMemoryGraph bridge) #1492: older controller availability/integration issue; the current static synthesizer is available and executes.
• ruflo-knowledge-graph: kg-extract over-counts type imports; kg-traverse references disabled/mis-wired controllers (3.7.0-alpha.69) #2049: previous knowledge-graph skill/consumer wiring issue. This report is the current adapter's input contract, not a dispute of its completed skill corrections.
• @claude-flow/memory registry references 8 agentdb exports that no installable agentdb provides — native controllers silently dead on every fresh install #2977: missing native exports / the explicitly marked hierarchical consolidation stub. This report does not propose implementing that unavailable consolidator; SONA is a separate path.

A small reversible, exact-anchor downstream patch will follow with implementation and regression evidence.