Import an MCP Toolbox configuration into a semantic model

[libei]

Adds kcmd toolbox import <file|dir>: reads an MCP Toolbox for Databases
configuration and reconstructs the semantic model its tools imply.

Why it is not a one-to-one mapping

A Toolbox tools.yaml is a list of callable things. It has verbs and no
nouns — in the published quickstart the word hotels never appears as a thing,
only inside SQL, five times, as a table somebody happened to query. A semantic
model is a list of things that are true, with the callable surface derived
from them.

So the import cannot map verb to verb. It parses every statement and
reconstructs the nouns out of the SQL inside the verbs:

in the configuration	in the model
a table named in any statement	an entity, bound to that table
a column attributed to a table	a field, expression = the column
a join equality in a SELECT	a relationship, oriented by the foreign key
an INSERT/UPDATE/DELETE tool	an action with a sql executor
the tool's parameters	action parameters, rebound to @name, carrying description / required / default
the DML target, verb and columns	the action's affects
the source's coordinates	the entity source, and the model's deployment_target one level up
a SELECT tool	nothing

The last row is the point: a lookup over an entity is generated from the model,
not stored in it. The AlloyDB quickstart's four tools import to two actions and
three lookups nobody wrote, and kcmd agent tools lists all five.

Notes on the mapping

affects is derived, not trusted. The Toolbox annotations say only
destructive; the statement says which table, which verb and which columns,
and id is excluded because the statement reads it to find the row rather than
changing it.

Positional placeholders are rebound by index, not by order of appearance —
the quickstart's SET checkin_date = CAST($2 ...), checkout_date = CAST($3 ...) WHERE id = $1 against parameters hotel_id, checkin_date, checkout_date would
otherwise bind the row's identity to a date.

Join direction is decided from evidence in the text: a column naming the other
table is the foreign-key side. Where nothing in the text decides it, the written
order is kept and a warning says so.

Parameter metadata rides on what #442 added, and the description is the string
that mattered — it is what the agent reads to decide what to put in the
argument, and it now reaches the derived tool listing verbatim. Toolbox's
required is nil-defaults-to-true, the opposite of how an absent flag
usually reads, so it is written out explicitly rather than passed on as an
ambiguity; a defaulted parameter states its default and not its necessity,
since push rejects required: true beside a default.

deployment_target is written from the source coordinates, which is what
separates a model that loads from a model that runs — kcmd asks for one before
it offers an agent any tool. A configuration naming two databases gets none and
a warning naming both.

What a configuration cannot say

Reported as warnings rather than guessed at, because each is a real difference
between the formats and not an unfinished corner:

• no primary key — nothing states one; join columns are named as candidates
  and no key is written, since a wrong key is worse than an absent one
• no value restriction on a parameter — allowedValues is a rule about the
  argument, and a rule belongs in a constraint guarding the action, which the
  warning says; likewise a parameter filled from a verified claim imports as an
  ordinary one, which is a real loosening and is called out
• no toolsets — no construct for a named subset of the surface
• no templated statements — templateParameters are interpolated before the
  statement is prepared, so the blast radius cannot be declared ahead of the
  call; those tools are skipped
• nothing non-SQL — a tool wrapping an API has no statement to read an
  entity out of

A non-Google source (plain postgres, mysql) produces a model that loads,
validates and publishes to Knowledge Catalog but binds to no store. The import
says so at import time rather than leaving it to the first call.

Tests

Both published Toolbox quickstart configurations are the fixtures, unedited,
so the tests say what a real configuration yields rather than what a
purpose-written one does. 30 tests cover the converter and assert the output
loads via loadModels and passes validatePushRequirements; 1003 pass across
the semantic suite.

osi_schema.test.ts skips the new fixture directory — those are converter input
in a foreign format, the way the OWL inputs are Turtle.