feat(context): mode-aware axme_decisions/axme_memories + active KB triggers

[George-iam]

Summary

• axme_decisions and axme_memories now adapt output to config.context.mode — full mode unchanged, search mode returns a catalog (id/slug + title + ≤200-char description). Closes a v0.5.0 regression where these tools silently loaded full bodies in search mode, erasing the ~10× token saving.
• axme_oracle unchanged — always full bodies (stack/structure/patterns/glossary are connected docs, not catalog entries).
• buildSearchModeInstructions gained an "Active KB usage" block with concrete trigger predicates so the agent calls axme_search_kb / axme_get_* proactively (was: generic "useful for fuzzy lookups").

Why

In a session started today on this repo with context.mode: search, the model loaded the catalog via axme_context (page 2) and then immediately called axme_decisions / axme_memories because the unified startup instruction asked for "in parallel" loads — the tools returned full bodies, defeating the whole point of search mode. User pointed it out and asked for a reliable, mode-aware fix instead of additional prompt text.

Design (per discussion in this session):

• Server-side gate inside the tools themselves (single source of truth = config). Instructions stay unified across modes.
• Tools detect mode and produce the right shape. No bans, no agent-side branching.
• Active-usage triggers in the search-mode instruction make the agent reach for search proactively when the situation matches a recognizable predicate.

Test plan

• npm test — 542/542 (was 536, +6 new in test/context.test.ts)
• npx tsc --noEmit clean
• npm run build clean
• Manual: open a session with context.mode: search, call axme_decisions → catalog only; flip to full → bodies. Done in a follow-up window after merge.

Notes

CHANGELOG entry added under [Unreleased] (no version bump in this PR — bundle into next release per D-045 batch policy).

Includes a small KB cleanup as part of the same effort:

• Q-003 answered (auditor flagged 102 MB → 773 MB transformers.js size discrepancy).
• Stale memory transformers-js-install-size-is-102mb deleted (B-005 shipped, lazy-install pattern now in code, memory was historical noise).

🤖 Generated with Claude Code