#326: place-name search works — point facets_url at rebuilt sample_facets_v4

[rdhyee]

What this does

Fixes the missing-results half of #326: free-text search couldn't find "Axial seamount summit caldera" even though the samples table displays it in the Place column.

Root cause

Search scores label/description/place_name over facets_url — but the deployed isamples_202608_sample_facets_v3.parquet predates the #311 builder fix (merged in #318), so its place_name is NULL for all 6,026,242 rows. The table's Place column reads samples_map_lite_v3, which DID get the fix in #319. Result: the UI shows places that search can literally never match. (This divergence-between-sibling-derived-files failure mode is also being fed into a broader pipeline↔UI coherence review.)

Data change (mirrors #319 exactly)

Rebuilt sample_facets from the live prod isamples_202608_wide.parquet (downloaded fresh) with the current post-#311 scripts/build_frontend_derived.py, uploaded to R2 as isamples_202608_sample_facets_v4.parquet — a NEW filename, never an overwrite (immutable-cache rule). Validation:

• 6,026,242 rows — identical universe
• place_name non-null 2,263,648 (37.6%) — identical to #311: deploy the corrected samples_map_lite (real Place/Date data) #319's lite rebuild
• ZERO rows differ from _v3 on any non-place column (FULL OUTER JOIN over pid/source/material/context/object_type/label/description)
• exact-phrase "axial seamount summit caldera" → 284, exactly matching lite_v3
• explorer.qmd diff is one line: facets_url → _v4

Verified on fork preview (fresh isolated browser context)

• Andrea's exact search → 284 results, table rows show "Axial Seamount summit caldera" (incl. IGSN:321000001), 0 console errors
• Her second term ("seamount summit caldera") also returns 284 now

The "hang" half of #326 (not fixed here, honestly scoped)

Measured on the preview: cold search ~9.3s; a follow-up search ~24s even warm (queued behind post-search surface refreshes; the #190 full-HTTP-read fallback pulls the 69.4 MB facets file on cold load, which on a slow connection reads as "stuck"). This latency is the documented #168 "interim recall fix" tradeoff and predates this PR — the real cure is the FTS track (#169–#172), for which Andrea's report is fresh evidence. Follow-up comment will be posted on #326.

Codex review

LGTM, no blockers ("No consumer breakage: _v4 preserves the PID universe, uniqueness, schema, and every non-place value; consumers only gain populated place_name"). Non-blocking note: comments mentioning facets_v3 generically — deliberately untouched to keep this diff one line.

Fixes the results half of #326.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AjnkWb4HpuLeDbYfzmY3X9

[rdhyee]

Codex 5.6 verdict (independent review): "approve for production. No branch-level blocker. The one-line change is correct, and the evidence is unusually strong: identical PID universe, identical non-place columns, matching place coverage, and matching phrase count… I would merge this fix rather than hold a user-visible correction hostage to the broader audit." (Noted non-blocker: deployment provenance — neither Git nor the browser can independently verify the URL identifies the validated artifact; that's the subject of the coherence review now underway.)

— rbotyee 🤖

[rdhyee]

CI finding during preview proof-of-life (Raymond asked for a fork-preview deploy + confirmation before each prod merge): the pre-deploy smoke's "pottery" search failed its 90s budget twice (~07:15), after the identical commit passed at 06:20. Diagnosis: the smoke cold-downloads the FULL facets parquet on CI (#190 fallback) before scanning, 90s was already documented as marginal for broad terms, and this PR's rebuild grows the file 62.5→69.4 MB (real place strings replacing NULLs). Boot checks were green both times; only the search-latency wait tripped.

Pushed 04806dc: budget 90s→150s with a comment explaining why — the gate is a liveness check (dead search wiring still fails fast via pageerror/fatal-console asserts), not a latency SLA; latency is tracked by #167/#190 and really fixed by FTS #169–#172.

Also noted for the coherence review: "smoke budgets calibrated against a specific artifact size" is another quiet pipeline↔UI coupling — a data rebuild can flip CI red (or worse, stay marginally green) without any code change.

— rbotyee 🤖