docs: add a first-collection tutorial teaching schema design

[adityamparikh]

What

Adds docs/tutorial.md — a hands-on walkthrough taking a reader from an empty Solr to a
designed, queryable collection, entirely through natural-language conversation with an
assistant. Linked from the README documentation index.

Why

The docs cover how to connect a client and what each tool does, but nothing explains
why field types decide which questions are answerable. That gap matters here more than in
a normal Solr context: the assistant will happily index anything schemaless, and the user has
no obvious signal that they've traded away faceting and sorting until a question quietly
fails.

So the tutorial indexes the same 61-document dataset twice — once letting Solr guess, once
choosing types on purpose — and puts the two side by side:

	schemaless	designed
Documents indexed	61 of 61	61 of 61
platform type	text_general	string + docValues
imdb_rating type	pdoubles (multi-valued)	pdouble
"breakdown by platform"	fails	Netflix 20, Amazon Prime Video 20, HBO Max 7, …

Same data, same question, same tool. The only difference is that someone decided what the
fields meant.

Verification

Every factual claim was checked end-to-end against Solr 9 in SolrCloud mode, driven through
the MCP server over STDIO — not written from what the tool signatures imply:

• the guessed types (text_general, pdoubles, plongs) and the array-wrapped documents
• all eight search examples and their result counts (1, 2, 20, 31, 15, status 33/28, top-rated
  Game of Thrones 9.2, combined 39)
• the facet buckets quoted in the text
• check-health reporting 61 docs, and get-collection-stats exposing the cache hit ratios
  the tutorial invites the reader to ask about

Known issues, documented rather than hidden

A reader following the steps will hit both of these, so the tutorial names them and links the
issues instead of quietly routing around them:

• search: ClassCastException when a facet returns zero buckets #182 — faceting a query that matches nothing throws
  ClassCastException: ArrayList cannot be cast to NamedList. This is exactly what the
  schemaless pass runs into.
• create-collection binds every collection to the shared _default configset, leaking schema between collections #183 — create-collection binds collections to the shared _default configset, so
  schemaless guesses leak into every later collection and add-fields then fails with
  Field 'x' already exists. This is why the tutorial resets Solr midway, and the reset is
  called out as a workaround rather than presented as normal practice.

If either is fixed, the corresponding section should be trimmed — they're marked clearly
enough to find.

Notes for review

• Uses the raw shows.json from src/test/resources/ via its GitHub URL. Referencing a test
  resource from user docs is a little irregular; happy to move the dataset somewhere more
  deliberate if preferred.
• Docker examples follow the existing docs/clients/*.md convention of a locally built image
  rather than a registry reference, since no official image is published yet.
• **/*.md is RAT-excluded and the existing docs/*.md carry no license header, so this
  follows suit.

[adityamparikh]

Updated to assume #185 is merged — this PR now depends on it.

With the empty-facet fix in place, Step 2 no longer throws. The schemaless facet returns:

{ "numFound": 61, "documents": [], "facets": { "platform": {} } }

Which is a better illustration of the point than the exception was. A stack trace reads as
"something is broken and it isn't your fault." A query that succeeds against 61 documents and
returns an empty breakdown is exactly the trap schemaless sets: nothing tells you the data is
fine and the field type is the problem.

Changes:

• Step 2 rewritten around the empty result rather than the crash
• the #182 entry dropped from Known issues (#183 and the sortClauses note remain)
• the platform row sharpened with a concrete consequence — searching platform:prime
  matches all 20 Amazon Prime Video shows, so the category has stopped being one value

Re-verified against a build of the #185 branch, driven over the MCP protocol: the empty facet,
the 20-hit token match, the three guessed types (text_general / pdoubles / plongs), and
the array-wrapped documents.

If #185 is rejected or reworked, say so and I'll restore the previous wording — the earlier
revision documented the exception and is still in this branch's history.

[adityamparikh]

Correction to my earlier comment: this PR's dependency is #175, not #185. #185 was a duplicate of #175 and has been closed; #175 is the PR that makes Step 2 return an empty facet rather than throwing. Nothing in the tutorial text changes — only the PR number this depends on.