OpenTender: identifier-first dispatch + name-equivalence gate (fixes #29)

[Nitjsefnie]

Implements #29 — registration-number dispatch with the name path gated on name equivalence, so Orange can never again attach Red-Orange e.U.'s furniture tenders to Orange S.A.

What's in it

• fetch_by_registration(country, registration_number) — indexed body_ids.id_value query, country-scoped, joined back to tenders, restricted to your allowlist (ORGANIZATION_ID, TRADE_REGISTER, HEADER_ICO, TAX_ID — not SOURCE_ID/BVD_ID/ETALON_ID/VAT).
• Lookup wiring, mirroring openaleph — this is a deliberate deviation from the issue's LookupDeriver wording, because that machinery is structurally single-dict-result (_collect_registry_sources requires a _bh_<id>() builder and _build_result_hit builds exactly one hit from one dict); a multi-hit registration+name source can't be expressed there without rewriting the pipeline's result handling. openaleph is your existing precedent for exactly this adapter class, so _opentender_strategies sits beside _openaleph_strategies with the same registry-gate/dedup shape, identifier-first (derived national IDs for the 8 built countries: at_fn, cz_ico, dk_cvr, ee_registry_code, fi_ytunnus, siren, lt_code, lv_regcode), name as fallback. If you'd rather the country→key map live on the adapter class, that's a trivial move — say so.
• Name gate — your option (b): _tender_bears_name, mirroring fetch_by_name's _bears_name/_normalise_name gating from openaleph (c95600f). No relevance thresholds, per OpenAleph: free-text name fallback produces false-positive hits for generic-name entities #21.
• One load-bearing fix to a pre-existing bug, called out rather than slipped in: _db_search_impl's fallback tokenized by whitespace-split, so any name with internal punctuation (S.A., e.U., A/S, s.r.o.) became an invalid FTS5 query (S.A.* → syntax error) and silently returned nothing — which would have made the name gate dead for real company names. Now tokenized FTS-safe (\w+). Verified old-vs-new over adversarial name shapes (hyphens, digits, diacritics): everything that worked before still returns the same rows; punctuated names go from silent-empty/error to working. This also improves the existing /search.
• NOT included: the REGISTRY flip. Your CLAUDE.md registration checklist requires the frontend count copy, README/docs/social updates, and the OKF bundle regen in the same commit (else the okf drift job goes red) — that's deliberately left as your registration step. What remains: add "opentender": OpenTenderAdapter() to REGISTRY, drop it from _DELIBERATELY_UNREGISTERED in test_sources.py, and run the checklist. Graceful DB-absence is proven either way (no OPENTENDER_DB_FILE → adapter methods return [], strategies return [], nothing crashes — tested for both the registered and unregistered states, so flipping is safe even before the artifact is hosted).

Tests (17 new; suite 2358 → 2375, 0 failed)

Fixture DBs built through the extractor's own _DDL/_insert_tender. Your real examples are reproduced as fixtures verbatim — SIREN 380129866, bodies named Red-Orange e.U. and Orange controls s.r.o. rejected while a genuine Orange S.A. passes the gate — plus the id_type exclusion, cross-country non-match, graceful absence, and two end-to-end pipeline-membership pins (an opentender hit must survive the _lookup_pipeline consume-loop and the /lookup-source retry path; each test fails if "opentender" is dropped from its list-result tuple). Both load-bearing behaviors are mutation-verified: deleting the id_type allowlist or bypassing the name gate turns the exact corresponding test red.

What this cannot verify (your artifact-side spot-check)

The 5 GB artifact isn't reachable here, so the per-country derived-ID ↔ body_ids.id_value format match is confirmed only for FR (siren, from your own worked example). If e.g. at_fn's normalized form differs from how DIGIWHIST stores Austrian registration numbers, the AT path would return nothing on real data. Worth one real subject per non-FR country against the artifact before registering.

---

Generated by Claude Fable 5 (brief, review), Claude Opus 4.8 (implementation)

[Nitjsefnie]

Closing in favor of #74. Since this was opened, main gained guard tests (test_adapter_is_registered / test_warm_opentender_db_noop_when_not_registered) asserting opentender stays OUT of the registry until the artifact-slimming block in #29 is resolved — this PR's router+registry wiring contradicts them, while #74 lands the identifier-first primitive in the scope #29 currently allows. The wiring half can return as a small follow-up once the artifact issue lands.