docs(claims): make every engine count state what it counts, and stop two thirteens colliding (#424)

[cevheri]

The Phase 0 wrap-up item in #424 asks that the published engine claims agree. This is that sweep,
measured surface by surface rather than assumed.

The collision this repo created for itself

Taking the wire-compatible relatives from nine to thirteen in #442 put a second, disjoint
thirteen directly beside the thirteen shipped drivers — in both README.md and DOCKERHUB.md,
separated only by the word "more". A skimmer reading the bolded lead-ins saw "Thirteen engines" and
then "Thirteen more engines" and had no cue they were different sets.

Both now name their denominator, and the README states what the two make together: thirteen
drivers reaching twenty-six named engines.

DOCKERHUB.md had the same fault twice:

• "Thirteen engines" sat above a table with fourteen rows — the fourteenth is the embedded
  LibreDB store, so a reader who counts gets a different number than the one stated.
• The relatives table below stated no number at all, and renders thirteen engines in ten
  rows because engines that behave identically share a row. A reader tallying rows got ten.

Stale copy, each verified against the code

Surface	Was	Now
charts/libredb-studio/README.md	seven engines	thirteen — Chart.yaml's own enumeration
operator/helm-charts/.../README.md	seven engines	same (byte-identical copy)
operator CSV keywords	seven engines	thirteen
deploy/rancher/E2E_VALIDATION_TASK.md	"Chart.yaml still names seven engines"	Chart.yaml names thirteen
packaging/winget/...locale.en-US.yaml.tmpl	header thirteen, body seven	body agrees

The chart README one matters most: that string is what ArtifactHub renders, so it undersold the
product to exactly the reader who checks a claim by reading it.

The operator keywords left Couchbase, ClickHouse, Druid, Elasticsearch, OpenSearch and Trino
invisible to OperatorHub search. Added to operator/config and propagated with
make -C operator bundle — the bundle diff is the six keyword lines and nothing else, createdAt
untouched.

What was deliberately NOT changed

The Rancher catalog prose still says ten engines. An audit flagged five sites there as wrong.
They are not: that wording is version-scoped and frozen on purpose — the catalog page describes a
released version and stays at ten until a tag carries the three new engines — and
CATALOG_LISTING.md:15-37 says so in its own accuracy gate. Changing it would have undone a
decision, so the only edit in that area is the E2E task file, which described a mismatch that no
longer exists.

Two relatives named in the listings

Under a rule the owner settled: a relative is named only where its caveats are cosmetic (it
reports a different version string). A marketplace listing is read during buyer diligence, so
anything with a data-correctness caveat stays out.

• Added: TiDB (mysql family), FerretDB (mongodb family) — joining MariaDB and the three Redis
  forks in the flatpak/flatpark metainfo and the Azure description.
• Excluded despite being full tier: Citus, TimescaleDB, YugabyteDB — all three report row
  counts that are wrong or zero rather than missing.
• Excluded on tier: StarRocks, CockroachDB, Materialize, RisingWave.

The Azure bullet still tallies to thirteen slots, each relative folded into its driver's slot, so
its "thirteen engines" headline stays honest.

Chart version 0.1.40 -> 0.1.41

charts/libredb-studio/README.md is a packaged chart file and 0.1.40 is already published
(confirmed against the live index), so #167 requires the bump. appVersion stays 0.12.0. The
operator's chart copy moves with it — two independent guards enforce byte-identity between
charts/libredb-studio/ and operator/helm-charts/libredb-studio/.

B53: what this PR deliberately did not fix

Ten sites in docs/AGENT.md, docs/AGENT_DATA_FLOW.md and src/lib/agent/ say the composed
grounding path leaves "the other nine engines" to their providers. Nine was correct when the
factory built eleven type-ids. It now builds fourteen.

The replacement is not the obvious one, which is why this is recorded rather than patched. The
thirteen external engines give eleven; the original basis — every type-id, embedded libredb
included — gives twelve. The sites never say which they count, and that ambiguity is the
defect. B49 complicates it further: a libredb connection can never be grounded at all, so
"twelve" is arithmetically faithful but describes a path one of the twelve cannot complete. Picking
a digit silently would repeat exactly the fault this PR exists to fix.

Gates

format · lint (0 errors) · typecheck · knip · test (all 30 groups) · build — green.
Plus the guards this change touches, each run directly: chart:check strict
(chart 0.1.41 / appVersion 0.12.0 in sync), readme:check
(13 engines and the install commands match README.md in README_zh.md, README_ja.md),
helm lint --strict (0 failed), the operator bundle freshness gate reproduced as CI runs it
against the commit (exit=0), the two chart trees confirmed byte-identical, and
appstreamcli validate on both metainfo files.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud