fix: add amnesia mode and x-lore-no-store to prevent QA contamination in recall

[BYK]

Summary

Adds two mechanisms to suppress temporal storage and background work without affecting Lore's read path (LTM injection, recall tool, gradient transform):

1. /lore:amnesia:on|off — session-scoped toggle via slash command
2. x-lore-no-store — per-request header for programmatic use

Also generalizes /lore:* slash command handling so unknown commands get a helpful error instead of being forwarded upstream.

Problem

The eval's QA questions were sent through the gateway as new sessions, creating temporal messages and distillations that contaminated recall search results. When QA question N+1 invoked recall, it found QA session N's responses (which literally contained the question terms) instead of the actual replay session content.

Changes

Gateway (pipeline.ts)

• handleLoreSlashCommand(): Generic dispatcher for all /lore:* commands — routes to warm/curate/amnesia handlers, returns error for unknown commands
• handleAmnesiaSlashCommand(): Toggles sessionState.amnesia flag
• postResponse(): Gates temporal storage and background work on !noStore where noStore = sessionState.amnesia || req.rawHeaders['x-lore-no-store'] === 'true'
• Refactored warmup handler to use shared slashResponse() helper

Types (translate/types.ts)

• Added amnesia?: boolean to SessionState

Eval (harness.ts)

• QA questions sent with x-lore-no-store: true header

Verification

• 1752 pass, 0 fail across all tests
• Typecheck clean across all 4 packages