docs: add SEARCH_INDEX_V1.md (closes #169)

[rdhyee]

Summary

The v1 search-substrate contract for the Explorer FTS work. Doc-only PR; closes #169.

• Sample-centric document projection with virtual fields (`sample.label`, `sample.description`, `sample.place_name`, `concept.label`).
• Concept labels in v1 minimum — dereferenced from `material` / `context` / `object_type` URIs via `vocab_labels.parquet`. Coverage data made this load-bearing: `description` is only ~24% populated on MaterialSampleRecord, but facet URIs are near-universal.
• Tokenizer: lowercase + NFKC + diacritic strip + whitespace split + length filter; no stemming; no stopwords (build-time).
• Query-time policy distinct from build-time tokenizer: drop English stopwords from the bag-of-words AND so `pottery from Cyprus` doesn't fail on `from`.
• BM25 with precomputed DF + doc_len, fixed `k1=1.2`, `b=0.75`. Field weights live in query code, not substrate data.
• Hash-by-token partition with a 5 MB per-shard cap; high-frequency tokens sub-shard by `hash(pid)`.
• Budgets are contract: cold ≤ 2 s, warm ≤ 500 ms, filter-composed ≤ 3 s, ≤ 5 MB cold + ≤ 1 MB warm. Quality thresholds calibrate after Explorer FTS Track 1a: Browser perf-smoke baseline #167 baseline lands.
• Hard quality gate: zero concept-only or stopword-heavy queries return empty; all hard-fail invariants in Explorer FTS Track 5: GO/NO-GO decision gate #172 must pass.
• `build_stats.json` artifact requirement — prevents the contract from drifting from the builder (Explorer FTS Track 3: Offline index builder + tokenizer regression set #170 §6).
• v1 ≠ destination: §1 names every v1.5 / v2 expansion vector, schema admits content-only growth, no migration.

Companion to `EXPLORER_STATE.md` (PR #166): that doc pins the UI-surface contract (option C — side-panel + result-pin overlay); this doc pins the backend contract. The two are intentionally orthogonal.

Test plan

• Review §1 v1 / v1.5 / v2 tiering against `query-spec.qmd:213-221` Solr surface
• Confirm field weights in §5 (label 3.0, concept 2.5, place 2.0, description 1.0) match intuition
• Confirm partition shape in §6 (hash-by-token, 5 MB cap, sub-shard rule for high-DF tokens)
• Confirm budget table in §7 is what we want as contract
• Confirm quality-gate shape in §9 (top-K vs hand-labeled + DuckDB FTS oracle, hard-fails)
• Confirm `build_stats.json` schema in §10 carries the right empirical signals

Closes #169. Refs #165, #166, #170, #171, #172, #174.

🤖 Generated with Claude Code

[rdhyee]

Review result: no blocking issues found. The SEARCH_INDEX_V1 contract has a coherent v1/v1.5/v2 split, separates build-time tokenization from query-time stopword policy, and includes the build_stats requirement needed to keep implementation honest.