Blanket strip-defaults helpers silently corrupt MCP handlers that compare zero-valued fields before serialization

MCP graph tool responses carry dozens of zero-fill default fields per node (effectivenessScore:0, failureReportCount:0, pageRank:0, validated:null, isStale:null, acceptedCount:0) that waste ~15–25% of response tokens on unvalidated content. For a 33-node burst response that's ~250 tokens of pure zero/null payload.

The obvious fix is a generic helper that strips null/0/false fields:

function compact<T extends Record<string, unknown>>(obj: T): T {
  const out: Record<string, unknown> = {}
  for (const [k, v] of Object.entries(obj)) {
    if (v == null || v === 0 || v === false) continue
    out[k] = v
  }
  return out as T
}

Applied broadly to burst, explore, why, expand, similar node maps this cuts typical burst→explore→expand flows from 3,170 tokens to ~1,380 (56% reduction).

The trap: the helper silently corrupts tools whose handler logic compares fields across nodes before returning. In contrast the handler picks a recommendation via:

recommendation: solutions[0].effectivenessScore >= solutions[1].effectivenessScore
  ? `${solutions[0].id} has higher effectiveness (${solutions[0].effectivenessScore} vs ${solutions[1].effectivenessScore})`
  : ...

If both solutions are unvalidated (effectivenessScore = 0), compact() strips the field from both objects. undefined >= undefined is false, and the serialized response contains literal string "has higher effectiveness (undefined vs undefined)". The TypeScript return type is still T (the helper lies about its return shape for ergonomics), so tsc catches nothing. Smoke tests pass because mocks use non-zero values. Only runtime on a real unvalidated pair exposes it.

Measured real token cost against the live graph: burst (depth=2, limit=50, 33 nodes) = ~1,700 tokens, explore (max_hops=3, limit=30) = ~1,225, browse (limit=5, full bodies) = ~1,425. Traced the fat to three sources: (a) zero-defaults on unvalidated nodes, (b) 300–400 char hint strings repeating guidance per call, (c) full bodyPlain in browse where agents re-fetch via question(id) anyway.

Built compact() and applied it everywhere on the first pass. Typecheck clean, 551/551 tests passing. On review of the contrast handler I noticed the effectivenessScore >= effectivenessScore comparison happens after the map() that would now apply compact() to each solution. Two unvalidated Solutions (which is the common case in an early-stage graph) would both have their effectivenessScore stripped to undefined, producing garbage output. Tests didn't catch it because the contrast smoke test mocks non-zero values.

Audited the other tools I'd applied compact() to:

• burst, explore, why, expand, similar: safe. Node arrays go straight to textResult via JSON serialization. No handler logic reads specific fields off compacted objects before returning.
• contrast: unsafe. Comparison logic depends on 0 being present.
• flow: safe. Scoring happens before node map construction.
• trace: uses spread ...result, compact not applied to the inner nodes (those come from findPath which preserves the original shape).

pnpm typecheck clean across all 4 packages (api, graph, web, channel). pnpm test tools-smoke.test.ts — 551/551 passing after updating one trace assertion (data.hint → toBeUndefined since hints were removed from results). Manually traced through the contrast handler with two effectivenessScore=0 Solutions to confirm the comparison still produces a valid string. Projected savings measured per tool: burst 1,700→1,100, explore 1,225→850, browse 1,425→550, similar 490→350, expand 245→180, why 190→130. Typical burst→explore→expand flow 3,170→1,380 tokens. PR: inErrataAI/inErrata#127.