fix(eval): build realistic Lore context for QA questions

[BYK]

Summary

Fixes the CM-1 eval to test Lore the way it actually works in production — with distillation prefix + raw tail window + LTM + recall, using separate budgets that match the gradient context manager.

Problem

The eval's Lore baseline created a fresh gateway session per QA question. This new session had:

• No distillation prefix — distillations are session-scoped (gradient.ts:691), and the QA session had a different ID
• No context-bound LTM — gated by isFirstTurn (pipeline.ts:2900)
• Only the recall tool — which returned compressed/truncated results

In real Lore usage, the gradient system allocates separate budgets: 25% of usable context for distillation prefix, 40% for raw tail window, with LTM injected separately. The eval was testing recall-in-isolation rather than the full Lore experience.

Fix

Adds buildLoreContext() to the eval harness that mimics the gradient context manager:

1. Separate budgets — distillation and raw tail don't compete for the same pool (matching config.ts defaults)
2. Usable context — subtracts output reserve (32K) and overhead (~8K) before computing budgets (matching gradient.ts logic)
3. Non-archived distillations only — archived = 0 filter matches what the gradient system includes in-context
4. No LTM duplication — LTM is left to the gateway's own system prompt injection
5. Built once per baseline — not per question (performance)

The recall tool is still available via the gateway for additional detail.

Results

Base CM-1 (~16K tokens):

Baseline	Before	After
Lore	1.9	4.8
Tail-window	4.9	4.9
Compaction	4.9	4.9

Inflated CM-1 (400K tokens) — Lore beats tail-window by +50%:

Baseline	Score
Lore	3.9
Tail-window	2.6
Compaction	4.1

Remaining gap vs compaction (3.9 vs 4.1) is due to distillation compression losses — tracked in #417.

Changes

packages/core/eval/harness.ts

• Add buildLoreContext(turns, contextWindow) — builds distillation + raw tail context with separate budgets matching gradient defaults
• Update askQuestionViaGateway() to accept and inject Lore context into the QA request
• Hoist context building outside the question loop (built once per baseline)
• Remove unused sessionHeaders parameter

Related to #410