explorer: interim recall fix — search description (#168 Direction A, knowingly slower)

[rdhyee]

Summary — interim recall fix, NOT the future FTS backend

`doSearch()` swapped from `samples_map_lite.parquet` to `sample_facets_v2.parquet` so live search now covers `label + description + place_name` instead of `label + place_name`. This trades latency for honest recall, knowingly violating the latency thresholds locked in #168.

The previous live search returned zero results for `pottery Cyprus` even though ~7,124 samples actually match (Cyprus appears only in description). The #168 honesty-fix landed in #176 documented this gap; this PR closes it functionally as an interim state while the proper substrate work in #169-#172 builds the real FTS that fixes both recall and latency.

Why this is shipping despite the threshold failure

• Explorer FTS Track 1b: Honesty fix for query-spec / live mismatch #168 locked: cold `pottery` ≤ 5 s, multi-term ≤ 6 s.
• Explorer FTS Track 1a: Browser perf-smoke baseline #167 baseline measured the swap at 12 s and 14.8 s respectively.
• Strictly per the threshold rule, Direction B (the doc fix in docs: query-spec honesty-fix for live search field set (#168 Direction B) #176) was the answer. That's what merged.
• But the threshold was set when we expected the swap to be roughly latency-neutral. It isn't. And "fast and broken" (current behavior on `pottery Cyprus`: 5 s, returns 0 results) is worse UX than "slow and useful" (this PR: 15 s, returns 50 results).
• A search bar that tells you `pottery Cyprus` is a good example query, then returns nothing for it, is broken. We made the placeholder example honest first; this PR makes the underlying behavior honest.

This is the kind of decision the threshold rule was a guess about, not an answer to. The data flipped the answer.

What this is NOT

• Not the FTS solution. Explorer FTS Track 2: search_index_v1 contract doc #169-Explorer FTS Track 5: GO/NO-GO decision gate #172 builds the real substrate: sample-centric document projection (label + description + place_name + dereferenced concept labels), hash-partitioned BM25 indexes, sub-second cold search, sub-500-ms warm. This PR's interim search goes away the day Explorer FTS Track 4: Browser query prototype + benchmark #171 lands.
• Not a reversal of the substrate decision. The state contract (PR docs: explorer state contract (closes #164) #166) and search-index contract (Explorer FTS Track 2: search_index_v1 contract doc #169 / PR docs: add SEARCH_INDEX_V1.md (closes #169) #175) stand. Track 6 (hosted-search) remains a contingency.
• Not a precedent for relaxing locked thresholds. Future locked thresholds remain locked unless similarly contested by data.

What changes

• `explorer.qmd` `doSearch()` SQL: CTE-then-keyed-join over `sample_facets_v2` (filter to top-50 first, then keyed JOIN to `samples_map_lite` for coordinates). Native DuckDB benchmark showed the naive LEFT JOIN form is 8× slower (4.2 s vs 0.5 s for `pottery`) and times out in browser; the CTE form is the only shippable shape.
• `explorer.qmd` search-help line: updated to "Searches labels, descriptions, and place names. First search can take 10-15 seconds while data loads; subsequent searches are faster." Forward-link to Explorer FTS Track 2: search_index_v1 contract doc #169.
• `explorer.qmd` placeholder example: `pottery Cyprus` (a query that now works).
• `query-spec.qmd`: §3.2 wording reflects the interim state with explicit "interim" framing; §5.1 binding describes the new query shape.

Cold-search caveat

The new `.search-help` UI line warns users that the first cold search can take 10-15 seconds. This is honest. The substrate work fixes it. Users who hit that latency see a "Searching..." indicator (existing behavior) plus the new explanatory line.

A future PR could add an explicit progress spinner or shimmer; out of scope here.

What this unblocks

• `pottery Cyprus`, `Çatalhöyük` (in description), `100%` (in description) — all flip 0 → 50 results.
• The substrate work (Explorer FTS Track 3: Offline index builder + tokenizer regression set #170, Explorer FTS Track 4: Browser query prototype + benchmark #171, Explorer FTS Track 5: GO/NO-GO decision gate #172) is not blocked by this; the interim and the substrate are independent code paths and the substrate replaces this when ready.

Test plan

• Open Explorer; verify `pottery Cyprus` now returns 50 results in ~10-15 s cold, ~3-5 s warm.
• Verify `Çatalhöyük` returns 50 results.
• Verify source filter still works (`pottery` × OPENCONTEXT only).
• Verify a result with no lat/lng still renders in side panel without crashing (LEFT JOIN can produce null coords for samples not in samples_map_lite).
• Verify the `?perf=1` panel shows the new search timings per the explorer: search perf-smoke baseline (#167) #173 instrumentation.
• Verify the search-help line is visible and reads honestly.

Closes #168 (functionally — the recall gap that PR #176 documented). Refs #165, #167, #169-#172, PR #166, PR #173, PR #175, PR #176.

🤖 Generated with Claude Code

[rdhyee]

Local-preview verification (per Codex review residual ask)

Ran against the post-merge local preview. Codex's two eyeball checks both clean:

`?perf=1` panel updates per search ✓

After running `pottery Cyprus` in the search bar, the panel grew from 8 boot-time rows to 9, and row 9 reads:

```
search #1: "pottery Cyprus" (50) 11.69 s
```

The per-search append code in `doSearch()` finally block (#173 round-2 fix) wires through. Subsequent searches add rows.

Search-help line reads cleanly ✓

DOM-extracted text:

Searches labels, descriptions, and place names. First search can take 10-15 seconds while data loads; subsequent searches are faster. Tracking issue: faster substrate FTS in progress.

Visible (helpVisible=true; bounding-box height > 0). Placeholder shows the new `pottery Cyprus` example.

Recall recovery confirmed ✓

`pottery Cyprus` returns 50 results (was 0 in the pre-#177 baseline). Camera flies to first result at 34.99, 33.71 — Cyprus coordinates.

Cold elapsed 11.69 s — matches the #167 baseline measurement (12 s for multi-term) within noise. The latency cost is real and was knowingly accepted in the PR framing.

Ready to merge.