feat(diversity): per-platform Layer-A state, exhaustion warnings, gro…

[PlevanTem]

…w-corpus.js

Three coordinated changes to make the corpus-direct diversity system persistent and infinitely extensible across batches:

1. Per-platform Layer-A usage state (run-corpus.js)

   • Default path now corpus_usage_{platform}.json (was corpus_usage.json)
   • Platform is resolved before usageStatePath so mobile and web never write into each other's state; aligns with the per-platform files already written by merge-subagent-retry.js

2. Exhaustion warnings (query_factory_v2.js)

   • pickLeastUsedTopics() emits [Layer-A WARN] with l2Key when count > pool_size (cycling is about to start)
   • buildCorpusPlan() emits a pre-call warning when all topics in an L2 have been used at least once

3. grow-corpus.js — new script for LLM-based topic pool expansion

   • Scans corpus_data[_web].json for L2s with <threshold fresh topics
   • Generates expand-by new topics via Claude CLI per L2
   • Jaccard dedup (threshold 0.4) prevents near-duplicate additions
   • Appends in-place to corpus file; dry-run mode available
   • Usage: node scripts/grow-corpus.js [--platform mobile|web] [--threshold 10] [--expand-by 20] [--dry-run]

4. README: three new sections (No-API mode, grow-corpus.js, merge-subagent-retry.js) + extended parameter table for run-corpus.js (--platform, --prep-only, --prep-batch, --score, --no-xlsx)

5. docs/index.html: Layer-A cards + i18n strings updated to reference per-platform state files and grow-corpus.js; both en/zh blobs synced