feat: improve recall tool description + add cross-session cue eval scenarios

[BYK]

Problem

When a user says things like "we had this thing from earlier sessions" at the start of a conversation (layer 0, no compression), the LLM does not use the recall tool. The current tool description frames everything around "trimmed context" — which isn't true at layer 0, so the LLM dismisses the need to search.

Solution

1. Recall Tool Description Rewrite

Rewrote RECALL_TOOL_DESCRIPTION with a dual-trigger structure that separates two distinct cases:

1. Cross-session references (always true) — explicit cue phrases like "last time", "we discussed", "earlier", "remember". Prior sessions are never in context.
2. Missing details (true when compressed) — file paths, decisions, preferences not visible in the current window.

This works at every gradient layer because cross-session content is never in context, regardless of compression state.

2. Eval Extension for Cross-Session Recall Cue Detection

Extended the eval suite to test whether the LLM uses recall when given conversational cross-session references:

• x-lore-recall-invoked response header — added to all non-streaming recall return paths in the gateway pipeline so the eval harness can detect recall usage (the gateway handles recall transparently — clients never see tool_use blocks)
• RECALL_TRIGGER scoring criterion — new judge criterion that scores whether the LLM appropriately used recall for cross-session references (1-5 scale)
• crossSessionCueRecall rubric — factual_accuracy (0.25), completeness (0.25), recall_trigger (0.3), temporal_attribution (0.2)
• recallInvoked metadata flow — propagated from askQuestionViaGateway() → judge() → EvalResult.metadata, conditionally included in the judge prompt only when the rubric has a recall_trigger criterion
• 8 new MSR-1 questions (msr1-q13 through msr1-q20) — same factual content as existing questions but phrased with natural conversational cues:
  • "Remember that auth bug we had?"
  • "We set up something for token refresh last time"
  • "Earlier we discussed why we went with JWT"
  • "What was that password hashing library we picked"
  • "remind me why" (PKCE)
  • "We changed the rate limiter key at some point"
  • "We built up the auth module over a couple of sessions"
  • "What were those callback URLs we configured"

Files Changed

File	Change
packages/core/src/recall.ts	Rewrite RECALL_TOOL_DESCRIPTION
packages/gateway/src/pipeline.ts	Add extraHeaders to nonStreamHttpResponse(), set x-lore-recall-invoked header
packages/core/eval/judge.ts	Add RECALL_TRIGGER criterion, crossSessionCueRecall rubric, metadata param to judge()
packages/core/eval/harness.ts	Read recall header, propagate recallInvoked through scoring pipeline
packages/core/eval/scenarios/multi-session-recall.ts	Add 8 new cross-session cue questions

Verification

• bun run typecheck — all 4 packages pass
• bun test — all 1630 tests pass
• All changes are backward compatible (new params are optional, existing questions unaffected)