pyoq-sql 1.0.2__py3-none-any.whl

pyoq_sql-1.0.2.dist-info/METADATA

@@ -0,0 +1,3050 @@
+Metadata-Version: 2.5
+Name: pyoq-sql
+Version: 1.0.2
+Summary: A fully typed, database-first SQL toolkit for Python.
+Project-URL: Homepage, https://teqpod.com/pyoq-sql/
+Project-URL: Documentation, https://teqpod.com/pyoq-sql/overview/
+Project-URL: Issues, https://github.com/Teqpod/pyoq-sql/issues
+Project-URL: Security, https://github.com/Teqpod/pyoq-sql/security/advisories/new
+License-Expression: MPL-2.0
+License-File: LICENSE
+Keywords: database,query-builder,sql,typing
+Classifier: Development Status :: 5 - Production/Stable
+Classifier: Framework :: AsyncIO
+Classifier: Framework :: Django
+Classifier: Framework :: FastAPI
+Classifier: Intended Audience :: Developers
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3 :: Only
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Programming Language :: Python :: 3.14
+Classifier: Topic :: Database
+Classifier: Typing :: Typed
+Requires-Python: >=3.11
+Provides-Extra: django
+Requires-Dist: django<7,>=4.2; extra == 'django'
+Provides-Extra: fastapi
+Requires-Dist: fastapi<1,>=0.115; extra == 'fastapi'
+Provides-Extra: mysql
+Requires-Dist: pymysql<2,>=1.1; extra == 'mysql'
+Provides-Extra: mysql-async
+Requires-Dist: asyncmy<1,>=0.2.14; extra == 'mysql-async'
+Requires-Dist: pymysql<2,>=1.1; extra == 'mysql-async'
+Provides-Extra: postgres
+Requires-Dist: psycopg<4,>=3.2; extra == 'postgres'
+Provides-Extra: sanic
+Requires-Dist: sanic<26,>=24.6; extra == 'sanic'
+Provides-Extra: tracing
+Requires-Dist: opentelemetry-api<2,>=1.25; extra == 'tracing'
+Description-Content-Type: text/markdown
+
+# PyOQ
+
+PyOQ is a fully typed, database-first SQL toolkit for Python. It is designed to
+provide explicit query construction, generated schema types, predictable
+execution, and efficient result hydration without hiding SQL behavior.
+
+SQLite, PostgreSQL and MySQL are supported, synchronously and asynchronously.
+Schema reflection generates row, key, insert, update, table, column, relation,
+enum, domain and routine types that hold under strict type checking. Relations
+are fetched by a plan that keeps the query count independent of the result,
+and Django, FastAPI and Sanic each run on the same core rather than on a copy
+of it.
+
+Every public behaviour is covered by tests at complete statement and branch
+coverage, run against real PostgreSQL and MySQL servers rather than recorded
+stand-ins.
+
+## Installation
+
+```console
+python -m pip install pyoq-sql
+```
+
+The distribution name is `pyoq-sql`; the Python import package is `pyoq`.
+
+Releases provide native platform wheels and a portable fallback wheel, so
+installation does not require Rust. Building from the source archive does
+require a supported Rust toolchain. When the portable wheel is used, select
+its runtime explicitly:
+
+```console
+PYOQ_RUNTIME=python python -m your_application
+```
+
+The fallback preserves behavior and typing but is outside the native
+performance contract. `PYOQ_RUNTIME=native` makes an unavailable native engine
+a startup error. The default selects the native engine when present and emits
+a runtime warning before using the fallback when it is absent.
+
+## Package check
+
+```python
+from pyoq import __version__
+
+print(__version__)
+```
+
+## Package architecture
+
+Implementation code is grouped by capability. Each feature package owns its
+public facade and private implementation details, while package-wide errors,
+naming policy, typing metadata, and the native extension remain at the root.
+
+```mermaid
+flowchart TD
+    Package[pyoq] --> CLI[cli]
+    Package --> Config[config]
+    Package --> Generation[generation]
+    Package --> Query[query]
+    Package --> Runtime[runtime]
+    Package --> Schema[schema]
+    Package --> Serving[serving]
+    Package --> Errors[errors]
+    Package --> Naming[naming]
+    Runtime --> Native[private native extension]
+```
+
+Applications should import feature APIs from their stable facades, such as
+`pyoq.config`, `pyoq.schema`, `pyoq.generation`, and `pyoq.runtime`. Nested
+modules separate component responsibilities and are not compatibility
+surfaces.
+
+## Configuration
+
+Configuration is read from `pyproject.toml`. Generated code always lives in a
+dedicated directory below the project root. Use a complete environment
+reference when a data source name contains credentials.
+
+```toml
+[tool.pyoq]
+codegen-directory = "generated/database"
+codegen-package = "application.database"
+selected-profile = "default"
+
+[tool.pyoq.profiles.default]
+dialect = "postgres"
+host = "localhost"
+port = 5432
+database = "${POSTGRES_DATABASE}"
+username = "${POSTGRES_USERNAME}"
+password = "${POSTGRES_PASSWORD}"
+minimum-pool-size = 1
+maximum-pool-size = 8
+pool-checkout-timeout = 3.0
+
+[tool.pyoq.profiles.testing]
+dialect = "sqlite"
+database-path = "database/testing.sqlite"
+```
+
+Open the selected profile through the fully typed lifecycle API:
+
+```python
+from pyoq import connect
+from application.database.tables import ACCOUNT
+
+
+with connect(profile="default") as database:
+    rows = database.select_from(ACCOUNT).limit(10).fetch_all()
+```
+
+Omit `profile` to use `selected-profile`. The context creates one process-owned
+pool from the profile limits and closes it on exit. A framework-free
+application can instead pass a data source directly with
+`connect(dsn, dialect=Dialect.MYSQL)`. An explicit `pool_policy` replaces the
+complete profile pool policy for one call.
+
+SQLite file paths use `database-path`. They are resolved relative to the project
+root, must remain inside it, and are opened read-only during inspection. An
+environment reference in the exact `${VARIABLE_NAME}` form may be used for any
+structured server or pool value. Any other string is literal. Server profiles
+may alternatively define one `dsn`, but cannot combine it with structured
+connection fields.
+
+`SecretValue` renders as `[REDACTED]`; access to its underlying value is always
+explicit. An explicit `Configuration` object takes precedence over the file,
+and the `selected_profile` argument can override profile selection. Unknown
+settings, missing profiles, invalid dialects, unavailable environment values,
+and paths outside the project root fail with exceptions from `pyoq.errors`.
+
+## What each dialect can do
+
+Every row here is one operation, not a family, because a family name hides
+which parts of it exist. Every entry was taken from a running server rather
+than from a specification.
+
+**refused** means PyOQ raises before the statement is sent, naming what the
+dialect has not got. It never emits SQL a database cannot run, and never
+quietly does something narrower than what was asked. **written by PyOQ** means
+the dialect has no such operation and PyOQ writes it out of what the dialect
+does have, so one word means one thing on all three.
+
+This table is generated from the capability matrix the source keeps. A
+contract compiles every row against every dialect, checks that each answers
+with the kind of value the row claims, and names the dialect-neutral contract
+that runs it against a real server, so a claim here that is not true stops the
+test suite.
+
+<!-- capability matrix -->
+
+| Operation | sqlite | postgres | mysql | reads back as | run by |
+| --- | --- | --- | --- | --- | --- |
+| window: row number | yes | yes | yes | numeric | `window_contract` |
+| window: rank | yes | yes | yes | numeric | `window_contract` |
+| window: dense rank | yes | yes | yes | numeric | `window_contract` |
+| window: percent rank | yes | yes | yes | numeric | `window_contract` |
+| window: cumulative distribution | yes | yes | yes | numeric | `window_contract` |
+| window: ntile | yes | yes | yes | numeric | `window_contract` |
+| window: lag | yes | yes | yes | numeric | `window_contract` |
+| window: lead | yes | yes | yes | numeric | `window_contract` |
+| window: first value | yes | yes | yes | numeric | `window_contract` |
+| window: last value | yes | yes | yes | numeric | `window_contract` |
+| window: nth value | yes | yes | yes | numeric | `window_contract` |
+| window: an aggregate over a window | yes | yes | yes | numeric | `window_contract` |
+| window: partition by | yes | yes | yes | numeric | `window_contract` |
+| window: order by | yes | yes | yes | numeric | `window_contract` |
+| window frame: rows | yes | yes | yes | numeric | `window_contract` |
+| window frame: range | yes | yes | yes | numeric | `window_contract` |
+| window frame: groups | yes | yes | refused | numeric | `window_contract` |
+| window frame: exclude | yes | yes | refused | numeric | `window_contract` |
+| window: declared by name | yes | yes | yes | numeric | `window_contract` |
+| json: read a value | yes | yes | yes | json | `json_contract` |
+| json: read text | yes | yes | yes | string | `json_contract` |
+| json: how long an array is | yes | yes | yes | numeric | `json_contract` |
+| json: whether a path is there | yes | yes | yes | boolean | `json_contract` |
+| json: whether it holds a value | refused | yes | yes | boolean | `json_contract` |
+| json: set | yes | yes | yes | json | `json_contract` |
+| json: insert | yes | written by PyOQ | yes | json | `json_contract` |
+| json: replace | yes | yes | yes | json | `json_contract` |
+| json: merge all the way down | yes | refused | yes | json | `json_contract` |
+| json: write over members, one level | refused | yes | refused | json | `json_contract` |
+| json: remove | yes | yes | yes | json | `json_contract` |
+| json: build an object | yes | yes | yes | json | `json_contract` |
+| json: build an array | yes | yes | yes | json | `json_contract` |
+| json: name the members | written by PyOQ | written by PyOQ | yes | json | `json_contract` |
+| json: gather rows into an array | yes | yes | yes | json | `json_contract` |
+| json: gather rows into an object | yes | yes | yes | json | `json_contract` |
+| json: read an array as rows | yes | yes | yes | string | `json_contract` |
+| array: holds a value | refused | yes | refused | boolean | `array_contract` |
+| array: holds a value nowhere | refused | yes | refused | boolean | `array_contract` |
+| array: holds every value | refused | yes | refused | boolean | `array_contract` |
+| array: is among | refused | yes | refused | boolean | `array_contract` |
+| array: shares a value | refused | yes | refused | boolean | `array_contract` |
+| array: append | refused | yes | refused | other | `array_contract` |
+| array: prepend | refused | yes | refused | other | `array_contract` |
+| array: concatenate | refused | yes | refused | other | `array_contract` |
+| array: without a value | refused | yes | refused | other | `array_contract` |
+| array: replace a value | refused | yes | refused | other | `array_contract` |
+| array: one element | refused | yes | refused | other | `array_contract` |
+| array: how many elements | refused | yes | refused | numeric | `array_contract` |
+| array: how many along a dimension | refused | yes | refused | numeric | `array_contract` |
+| array: how many dimensions | refused | yes | refused | numeric | `array_contract` |
+| array: built from values | refused | yes | refused | other | `array_contract` |
+| lock: for update | refused | yes | yes | numeric | `locking_contract` |
+| lock: for share | refused | yes | yes | numeric | `locking_contract` |
+| lock: fail rather than wait | refused | yes | yes | numeric | `locking_contract` |
+| lock: skip what is locked | refused | yes | yes | numeric | `locking_contract` |
+| lock: named tables | refused | yes | yes | numeric | `locking_contract` |
+| lock: named fields | refused | yes | yes | numeric | `locking_contract` |
+| query: lateral source | refused | yes | yes | numeric | `lateral_contract` |
+| query: recursive common table | yes | yes | yes | numeric | `recursion_contract` |
+| query: row value | yes | yes | yes | boolean | `row_value_contract` |
+| query: cast | yes | yes | yes | string | `cast_contract` |
+| query: nulls first or last | yes | yes | written by PyOQ | numeric | `null_order_contract` |
+| query: right join | refused | yes | yes | numeric | `joined_contract` |
+| query: full join | refused | yes | refused | numeric | `joined_contract` |
+| routine: run a stored procedure | refused | yes | yes | rows | `routine_contract` |
+| routine: ask a function by name | yes | yes | yes | numeric | `routine_contract` |
+| routine: more than one result set | refused | refused | refused | rows | `routine_contract` |
+
+<!-- capability matrix -->
+
+A JSON path is steps rather than text, because PostgreSQL reads another
+dialect's written path as a member of that name and answers null without
+complaining. PostgreSQL's own insert raises where a member already exists,
+while the other two leave what is there, so PyOQ asks first on PostgreSQL.
+
+Merging and concatenating are two operations, not one. `json_merge` is
+merge-patch: a member given as null is removed, and an object inside an object
+is merged rather than replaced. MySQL and SQLite do that themselves, and
+PostgreSQL has no merge-patch at all. `json_concat` is one level deep: a member
+is replaced whole and a null is stored as a null. PostgreSQL does that itself,
+and the other two have nothing that means it. Each is refused where the
+dialect does not have it, rather than quietly doing the other one.
+
+Only PostgreSQL has an array type. SQLite will accept `TEXT[]` as a type name
+with no array behind it, which is the kind of tolerance an emulation would be
+built on, so the other two refuse instead.
+
+SQLite locks a database rather than rows and keeps no stored procedures. A
+timed lock wait is in no supported dialect: `FOR UPDATE WAIT` is a syntax error
+on both PostgreSQL and MySQL. Null treatment on a window function is likewise
+in none of them: `IGNORE NULLS` is refused by all three servers.
+
+A procedure that answers with more than one result set is refused rather than
+read down to its first, because a dropped result set is a silent one.
+
+### Schema objects
+
+| Object | SQLite | PostgreSQL | MySQL |
+| --- | --- | --- | --- |
+| tables, views, keys, relations, indexes, checks | yes | yes | yes |
+| enums | none | reflected and generated | reflected per column |
+| domains | none | reflected and generated | none |
+| routines | none | reflected and generated | reflected with typed calls or typed rejection |
+
+Only PostgreSQL keeps overloaded routines; MySQL refuses a second routine of
+the same name, and SQLite keeps no routine catalog at all.
+
+A procedure answers through the parameters it writes. A generated PostgreSQL
+call asks the caller only for IN and INOUT values, supplies required OUT
+placeholders internally, and reads the answer back with the catalog's types and
+nullability. MySQL needs a session-variable adapter for OUT and INOUT values.
+Those routines retain a generated, typed signature that raises
+`UnsupportedQueryError` before any invalid SQL is sent.
+
+A generated call names the schema the catalog said keeps the routine. Without
+it a routine outside the search path resolves to nothing, and one whose name
+another schema shares resolves to the wrong one.
+
+## Supported dialects
+
+Every configured dialect resolves to a schema source, so inspection and
+generation work for all of them. SQLite is implemented end to end. PostgreSQL
+has a compiler, type mapping, schema reflection, and both synchronous and
+asynchronous drivers with pooling, transactions, savepoints, and streaming.
+MySQL has all of that too, so all three dialects are now complete vertical
+slices.
+
+Database drivers are optional extras, so a SQLite installation stays dependency
+free:
+
+```console
+python -m pip install "pyoq-sql[postgres]"
+python -m pip install "pyoq-sql[mysql]"
+python -m pip install "pyoq-sql[mysql-async]"
+```
+
+Without them, PostgreSQL and MySQL query construction and compilation still
+work because they are pure Python. Only connecting requires a driver, and its
+absence raises an explicit error naming the extra to install. PostgreSQL and
+PostgreSQL and MySQL profiles supply their connection string through `dsn`, and
+inspection opens the connection read-only.
+
+A MySQL connection string names one database, which is reflected as the
+snapshot's schema:
+
+```console
+mysql://user:password@host:3306/database
+```
+
+Query and fragment options are refused rather than silently ignored, so an
+unsupported setting can never be mistaken for an applied one.
+
+`schema_source_for()` resolves a dialect to its schema source, and
+`supported_dialects()` reports every dialect that has one. Command services
+route through the configured dialect at load time, so adding a dialect is a
+registry entry rather than a change to the command layer.
+
+The layers a dialect plugs into are deliberately neutral. `CompiledQuery`,
+`StatementCompiler`, `SchemaSource`, `QueryOperations`, `BulkPlan`, and
+`DatabaseCursor` carry no dialect assumptions, so a new database supplies a
+compiler, a schema source, and a cursor-backed executor without reimplementing
+result cardinality, bulk planning, or ordered multi-operation execution.
+
+Execution is shared too. `ConnectionPool` owns the lease lifecycle, size and
+timeout policy, idle reuse, invalidation, and close semantics; `RowStream` owns
+bounded batch buffering, iteration ownership, and deterministic cleanup; and
+`TransactionState` owns scope nesting, savepoint naming, owner-thread checks,
+and stream interaction. A dialect supplies how it resets and validates a
+connection, how it opens a cursor, and how it begins a transaction.
+
+SQL rendering is shared as well. `ExpressionRenderer`, `QueryRenderer`, and
+`WriteRenderer` own the standard SQL that every relational dialect writes the
+same way: projections, sources, joins, common tables, ordering, set operations,
+inserts, assignments, row scoping, conflict resolution, and returning. A dialect
+declares only where it genuinely differs, such as how it spells distinctness,
+how it extracts a date part, how it renders string predicates, which joins and
+capabilities it allows, and how it paginates.
+
+Parameter placeholders are part of that declaration. `ParameterStyle` covers
+`?`, `$1`, and `%s`, and the compilation context escapes literal percent signs
+in structural SQL when the driver style requires it.
+
+`PostgresCompiler` shows the shape of a dialect built this way. It renders the
+same query model as SQLite while spelling distinctness as `IS DISTINCT FROM`,
+extracting date parts with `EXTRACT`, rendering string predicates with
+`POSITION`, `LEFT`, and `RIGHT`, allowing RIGHT and FULL joins and a native
+`NULLS` clause, emitting a bare `OFFSET` without a placeholder limit, and
+budgeting at the 65535 parameter ceiling of the extended query protocol. It
+rejects catalog-qualified identifiers and adding two temporal values, because
+PostgreSQL has no operator for either.
+
+MySQL has a compiler, type mapping, schema reflection, and synchronous and
+asynchronous drivers. It shows how far a dialect can diverge while still using
+the shared renderers: identifiers are quoted with backticks, string
+concatenation renders as `CONCAT` because `||` means logical OR in MySQL,
+distinctness uses the null-safe `<=>` operator, a bare `OFFSET` is given the
+maximum row limit MySQL requires, explicit null placement is written by
+ordering on whether a value is null because MySQL has no `NULLS` clause, and
+upsert renders as a row alias with `ON DUPLICATE KEY UPDATE`. The row alias form requires
+MySQL 8.0.19 or later; earlier servers and MariaDB use a different spelling
+that PyOQ does not emit.
+
+Where MySQL cannot express something, it says so instead of emulating it.
+`RETURNING` is refused with a message pointing at a follow-up query. A conflict
+target is refused because MySQL infers the key. A conflict condition is refused
+because `ON DUPLICATE KEY UPDATE` has no `WHERE`. Do-nothing conflict
+resolution is refused rather than emulated with `INSERT IGNORE`, which would
+also swallow unrelated errors, or with a self-assignment, which would report a
+different affected row count.
+
+`postgres_type()` normalizes declared PostgreSQL type names into the shared
+schema model, including serial types, spelled-out variants such as
+`timestamp with time zone`, length and precision arguments, and array types
+with their element type. A declaration it cannot parse stays opaque rather than
+being guessed at.
+
+MySQL executes through `MySQLExecutor` over a `MySQLPool`, with the same typed
+operations SQLite and PostgreSQL use. Four dialect behaviors differ and are
+documented rather than hidden.
+
+An affected row count means matched rows. MySQL counts changed rows by default,
+so an update writing a column's existing value would report nothing, and PyOQ
+connects with the client flag that restores the meaning every other dialect
+has.
+
+An insert reports the first identifier it generated, not the last, because that
+is what MySQL reports for a multi-row insert.
+
+A statement timeout is enforced from the client. MySQL applies its own
+`max_execution_time` only to read-only SELECT statements, so PyOQ sets that
+and also runs a watchdog that asks the server to stop the statement once the
+deadline passes. Cancellation works the same way, because MySQL has no
+client-side cancel. Both open a short-lived second connection, and a statement
+carrying neither a timeout nor a cancellation token starts no watchdog and pays
+nothing.
+
+A deadlock rolls the whole transaction back on the server while leaving the
+connection healthy, and a later commit then succeeds having committed nothing.
+PyOQ refuses to report that commit and raises instead, so work that was
+discarded is never reported as saved. A failed statement that is not a deadlock
+leaves a MySQL transaction usable, which is the opposite of PostgreSQL and
+needs no savepoint to recover from.
+
+Isolation is set before a transaction starts rather than as part of starting
+it, and the default is MySQL's own repeatable read rather than a level PyOQ
+imposes. Streaming reads through an unbuffered cursor, and a transaction allows
+only one open stream at a time so that a second statement cannot silently
+discard the rows the first has not read.
+
+`MySQLAsyncExecutor` over a `MySQLAsyncPool` provides the same operations
+without blocking, and behaves identically on every point above. Three things are
+worth knowing about it.
+
+It has its own extra, because its driver is a compiled one:
+
+```console
+python -m pip install "pyoq-sql[mysql-async]"
+```
+
+The synchronous extra stays pure Python. The asynchronous driver is compiled
+because the read path is worth it: PyOQ adds only a few percent above its
+driver on a large read, so the driver is the cost, and a pure-Python driver
+measured roughly an order of magnitude slower end to end on the same query. A
+throughput contract holds the read path to that standard so it cannot quietly
+regress.
+
+Its connections are encrypted. The synchronous driver negotiates TLS on its
+own, and the asynchronous one is given the same settings so that choosing
+`async` never quietly downgrades the transport. Encryption is what lets a
+server's password check happen without a separate cryptography library, so a
+server with TLS switched off needs `PyMySQL[rsa]` installed for either driver.
+
+Its driver publishes only generated type stubs that leave most of its surface
+unknown, and raises its own exception hierarchy rather than the synchronous
+driver's. PyOQ therefore declares the shape it depends on and confines every
+reference to the driver to one module. Failure classification is shared, since
+every MySQL driver reports the same numeric codes while raising different
+exception types.
+
+`mysql_type()` normalizes declared MySQL type names, including width and
+precision arguments, `unsigned`, `signed`, and `zerofill` modifiers, and the
+inline `ENUM` and `SET` declarations that carry their values in the column type
+rather than in a named type.
+
+MySQL reflection reads the connected database through `information_schema`.
+Auto-increment columns are reported as identity columns, stored and virtual
+generated columns keep their generation expression, and column defaults are
+normalized into valid SQL so that a literal default such as an empty string is
+quoted while an expression default such as `json_object()` is not. Descending
+and expression index terms, composite foreign keys with their update and delete
+rules, and CHECK constraints are all carried across. The placeholder comment
+MySQL stores on every view is discarded rather than reported as a comment.
+
+PostgreSQL executes through `PostgresExecutor` over a `PostgresPool`, with the
+same typed operations SQLite uses. Two dialect behaviors differ and are not
+hidden. A statement timeout is enforced by the server through
+`statement_timeout` rather than a client-side interrupt, and a cancellation
+token is honored mid-statement by a watchdog that issues a server cancel.
+Streaming uses a server-side cursor inside its own transaction, which is what
+keeps memory bounded for large results.
+
+PostgreSQL also executes asynchronously through `PostgresAsyncExecutor` over a
+`PostgresAsyncPool`. The asynchronous surface mirrors the synchronous one
+method for method, including pooling, transactions, savepoints, server-side
+cursor streaming, server-enforced timeouts, and cancellation:
+
+```python
+async with PostgresAsyncPool(PostgresAsyncConnectionFactory(dsn)) as pool:
+    database = PostgresAsyncExecutor(pool)
+    async with database.transaction() as transaction:
+        await transaction.execute(statement)
+    async with database.stream(query) as rows:
+        async for row in rows:
+            ...
+```
+
+Bulk writes use PostgreSQL pipeline mode on both paths. The planner still
+splits a large insert into chunks that fit the parameter budget, but the chunks
+are sent without waiting for each result, which removes a round trip per chunk.
+This changes one behavior for the better and it is worth knowing: a pipelined
+batch is atomic, so a failure discards the whole batch rather than leaving
+earlier chunks applied. Sequential chunking keeps the earlier chunks, as it
+always did.
+
+Pipelining is governed by the `pipeline` compiler capability and by whether the
+installed libpq supports it. When either says no, chunk execution falls back to
+the sequential path with identical results.
+
+Ownership is per task rather than per thread, so a stream or transaction used
+from a different task fails explicitly instead of corrupting its state. Both
+paths are held to the same dialect-neutral execution contract, so behavior does
+not drift between them.
+
+The most important difference is transactional: in PostgreSQL a failed
+statement aborts the whole transaction, and every later statement in that
+transaction fails until it ends. Recovery is a savepoint taken before the
+statement that might fail:
+
+```python
+with database.transaction() as transaction:
+    with transaction.savepoint() as attempt:
+        attempt.execute(statement_that_may_conflict)
+    transaction.execute(next_statement)
+```
+
+PyOQ does not wrap every statement in an implicit savepoint to paper over this,
+because that would add a hidden round trip to every write.
+
+PostgreSQL reflection reads the system catalogs directly rather than
+`information_schema`, so it recovers identity and generated columns, array
+element types, per-column comments, foreign key referential actions, expression
+and partial indexes with their sort direction, check constraints, materialized
+views, and multiple schemas. Catalog rows are validated as they are read, so a
+row of unexpected shape fails as a schema error instead of producing a silently
+wrong model.
+
+## Schema model
+
+Database metadata is normalized into an immutable, dialect-neutral snapshot.
+Source identifiers retain their exact case, spacing, and Unicode content. The
+model does not apply Python naming rules or depend on database drivers and web
+frameworks.
+
+```mermaid
+flowchart TD
+    Snapshot[SchemaSnapshot] --> Catalog[Catalog]
+    Catalog --> Schema[Schema]
+    Schema --> Table[Table]
+    Schema --> View[View]
+    Schema --> Enum[EnumType]
+    Table --> Column[Column]
+    Table --> Key[Key]
+    Table --> Relation[Relation]
+    Table --> Index[Index]
+    Table --> Check[CheckConstraint]
+```
+
+Construct snapshots directly when implementing metadata sources:
+
+```python
+from pyoq.config import DatabaseDialect
+from pyoq.schema import (
+    Catalog,
+    Column,
+    Identifier,
+    Schema,
+    SchemaSnapshot,
+    SqlType,
+    Table,
+    TypeKind,
+)
+
+user_id = Column(
+    name=Identifier("id"),
+    data_type=SqlType(TypeKind.INTEGER, "INTEGER"),
+    nullable=False,
+)
+users = Table(name=Identifier("users"), columns=(user_id,))
+snapshot = SchemaSnapshot(
+    dialect=DatabaseDialect.SQLITE,
+    catalogs=(Catalog(None, (Schema(None, tables=(users,)),)),),
+)
+```
+
+`SchemaSnapshot.to_json()` produces compact, versioned, deterministic JSON with
+Unicode preserved. `SchemaSnapshot.from_json()` performs strict field and type
+validation before constructing model objects. The round trip preserves ordered
+columns, nullability, raw defaults, identity and computed values, SQL type
+details, qualified references, comments, indexes, checks, relations, enums,
+views, and dialect capabilities.
+
+Every model uses frozen slots. Duplicate identifiers, invalid type dimensions,
+misaligned foreign-key columns, unknown local key or index columns, conflicting
+generated-value metadata, and unsupported snapshot versions fail with typed
+schema exceptions. Qualified relation targets may remain outside a snapshot so
+metadata sources can represent intentionally partial introspection safely.
+
+## Relations
+
+A foreign key is one fact that can be read from either side, so PyOQ derives
+both. `derive_relations()` turns a schema snapshot into typed relations that
+carry direction, cardinality, and whether a to-one relation may resolve to
+nothing:
+
+```python
+from pyoq.relations import derive_relations
+
+for relation in derive_relations(snapshot):
+    print(relation.direction, relation.cardinality, relation.optional)
+```
+
+Cardinality and optionality are read from the schema rather than declared:
+
+- A foreign key names at most one row of the table it points at, so following
+  it forward is always to-one. It is optional when any of its columns is
+  nullable, because a foreign key with a null part references nothing at all.
+- Following it backwards is to-many, unless the referencing columns are
+  themselves unique, in which case it is a to-one that may be absent. A unique
+  key over part of the referencing columns is enough, since a subset that is
+  already unique makes the wider set unique too. This is what recognizes a
+  table whose primary key is also its foreign key as a one-to-one extension
+  rather than a collection.
+- A to-many relation is never optional. The absence of children is an empty
+  collection, not a missing value.
+
+A self-referencing foreign key produces both directions on the same table.
+
+`build_relation_graph()` indexes those relations by the table they start from,
+so a schema can be navigated in either direction:
+
+```python
+from pyoq.relations import build_relation_graph
+
+graph = build_relation_graph(snapshot)
+for relation in graph.to_many_from(graph.resolve(team_reference)):
+    print(relation.target.table.name.value)
+```
+
+Endpoints and lookups are resolved to the way the snapshot names a table, so a
+foreign key written against a bare name reaches the qualified table it means,
+and a caller need not know how the snapshot qualified anything. A bare name
+that two schemas both carry is left alone, because it names neither.
+
+A snapshot describes one database and a foreign key may point beyond it, so a
+target the snapshot does not describe is reported through
+`unresolved_targets` rather than refused. Such a table stays navigable, because
+the table that declared the key still knows its side of it; only asking for the
+missing table's definition fails, and it fails by name. A key naming a column
+that a table it did resolve does not have is refused outright, because that is
+corruption rather than partial coverage.
+
+Generation emits both directions of every key as typed constants, each
+carrying its cardinality and whether it may resolve to nothing:
+
+```python
+from generated.relations import (
+    EMPLOYEE_TEAM_TEAM_ID,
+    TEAM_EMPLOYEE_TEAM_ID_REVERSE,
+)
+
+EMPLOYEE_TEAM_TEAM_ID.to_one  # True: an employee has one team
+TEAM_EMPLOYEE_TEAM_ID_REVERSE.to_one  # False: a team has many employees
+```
+
+A relation is named after the table across the key and the columns that carry
+it, with the reverse direction saying so. Nothing in the name depends on where
+a relation falls in a list, so adding a key to one table cannot rename the
+relations already generated for another. A key that points at its own table
+produces both directions on that table, which is why the direction is always
+part of the name rather than only when it would otherwise collide. Where a
+constraint has a name of its own, that name is used instead.
+
+Two keys that no name can tell apart are reported as a naming collision rather
+than resolved by guessing, and naming the constraint in the database is the fix.
+
+### Fetch plans
+
+A fetch plan says which relations a query brings back and how, and it is read
+back before anything runs:
+
+```python
+from pyoq.relations import FetchPlan, FetchRequest, FetchStrategy
+
+plan = FetchPlan(
+    team_reference,
+    (FetchRequest(employees, FetchStrategy.NESTED, (FetchRequest(badge),)),),
+)
+print("\n".join(plan.describe()))
+```
+
+```console
+employee [] via nested
+  badge via auto
+```
+
+Four strategies are available. `joined` brings a relation back in the same
+query, `nested` brings a collection back as one correlated result, `select_in`
+issues one further query per level, and `auto` leaves the choice to be made
+against the dialect's capabilities when the query is planned, so one plan can
+take the best route each database offers.
+
+`resolve_fetch_plan()` turns that intent into a decision against what a dialect
+can actually do, and records why:
+
+```console
+posts ordered by id via select-in: this dialect cannot order inside an
+  aggregate, and this collection's order matters
+  users via joined: a to-one relation is one row of the join
+```
+
+A fetch request says what a collection is ordered by with `by_column()` and
+`by_expression()`, each of which takes a direction:
+
+```python
+from pyoq.relations import by_column, by_expression
+
+FetchRequest(
+    employees,
+    order_by=(by_expression("lower(name)"), by_column("id", descending=True)),
+)
+```
+
+A column is checked against the related table, because the plan can see the
+schema. An expression is named rather than carried, because a plan is built
+below the layer that writes queries and does not reach up into one. The name
+is what the plan reports when it explains itself, and naming it is enough to
+plan with: whether a collection is ordered at all is what decides how it can
+be fetched, not what it is ordered by. The expression itself is given to
+whatever carries the plan out.
+
+A to-one relation rides the join, because it is one row the join already
+carries. A collection is nested where the dialect can aggregate one, and falls
+back to select-in where it cannot, or where the collection's order matters and
+the dialect cannot order inside an aggregate. That last case is measured rather
+than assumed: SQLite from 3.44 and PostgreSQL both accept an order inside their
+aggregate, while MySQL rejects it outright, and MySQL discards an order given in
+a derived table instead of honouring it. An aggregate in the wrong order is
+worse than a second query, so PyOQ takes the second query.
+
+A plan that names a strategy the dialect cannot carry out is refused rather than
+quietly changed, because a caller who asked for an ordered nested collection and
+received an unordered one has no way to notice.
+
+`validate_fetch_plan()` checks a plan against the relation graph once, before
+any query runs, so a mistake is reported against the schema rather than part-way
+through a fetch. A relation the table does not have is refused, a nested
+relation is checked against its own table rather than the root, and fetching one
+relation twice at the same level is refused. The same relation may appear at
+different levels, because a key pointing back is a different fetch rather than a
+repeat, and that is also why plan depth is bounded.
+
+### Relation loading state
+
+A relation is loaded, known to be absent, or was never fetched. Rows are
+detached values, so reading one never reaches a database, which is what makes
+the third state necessary: without it a relation nobody asked for would be
+indistinguishable from one that resolved to nothing.
+
+```python
+team.employees.value  # the fetched rows
+person.manager.value  # None, when the key resolved to nothing
+person.badge.value  # raises: this relation was not fetched
+```
+
+Reading a relation nobody fetched raises rather than answering, because any
+answer would be a guess and no answer can be produced without I/O.
+
+Run the model against a reflected database with:
+
+```console
+python examples/relation_model.py
+```
+
+## Row identity
+
+A join repeats a parent row once per child, and an outer join invents a child
+made entirely of nulls. Both are answered by asking what identifies a row, so
+identity is decided before any row is built:
+
+```python
+from pyoq.hydration import IdentityMap, row_key_for
+
+key = row_key_for(table, reference, projection.position_of)
+identity = key.identify(result_row)  # None when the result holds no row here
+```
+
+A primary key identifies a table. Failing that, a unique key no part of which
+is nullable. A table offering neither is identified by everything projected
+from it, which is the only sound answer left: rows it cannot tell apart are
+rows nobody can tell apart. A null among the identifying values means the
+result holds no row there rather than a row whose identity happens to be null,
+which is how an absent child is recognized.
+
+Identity is by value however a driver spells it, so buffers and arrays compare
+by their contents. A value no dictionary can hold, or a key column the query
+does not select, is reported against the table it belongs to rather than
+surfacing as a bare type or lookup error from inside a fetch.
+
+`IdentityMap` builds each distinct row once. It lives for one fetch and is
+discarded with it, so a row it returns can never be a stale row from an earlier
+fetch.
+
+## Hydration
+
+A join hands back one row per combination, so the same parent arrives once per
+child and the same child can arrive under several parents. `hydrate()` reads a
+result back into rows that carry what the plan asked for:
+
+```python
+from pyoq.hydration import hydrate
+
+teams = hydrate(node, result_rows)
+```
+
+Rows come back in the order the result first mentioned them, a row the result
+mentions twice is built once, and children keep the order they first appeared
+in. A null among a child's identifying values means the result holds no child
+there, so an outer join yields an empty collection or an absent value rather
+than a row made of nulls.
+
+The result is read twice: once to group it and once to construct it. A frozen
+row cannot be handed its children after it exists, so its children have to be
+known before it is built.
+
+One table read at two places in a plan yields separate rows, because each
+carries different children and one row cannot hold both. Within one place, a
+child several parents share is one object.
+
+A to-one relation that returns several distinct rows for one parent is
+reported. The schema said those columns were unique and the result disagreed,
+and discarding rows would hide the disagreement.
+
+Throughput and memory are held to a contract:
+
+```console
+python -m benchmarks.hydration
+```
+
+## Query diagnostics
+
+A query issued once per row of a previous result is invisible from inside the
+loop that causes it. Recording what a scope executed makes it visible from
+outside:
+
+```python
+from pyoq.diagnostics import QueryObserver
+
+observer = QueryObserver()
+...
+for repeated in observer.repeated():
+    print(repeated.describe())
+```
+
+```console
+4 executions of SELECT id FROM person WHERE team_id = ? from app/teams.py:31 in members
+```
+
+One query is recognized across the many times it is executed. Its bound values
+are already placeholders, a list of any length collapses to one, a write chunked
+to fit a parameter limit collapses however many chunks it took, and the three
+placeholder styles reach the same shape, so one query compiled for three
+dialects is one shape. Placeholders that are not a list stay apart, so a query
+with two conditions never merges into one with a single condition.
+
+A report carries the first place outside PyOQ that issued the query, which is
+the caller's own code rather than the library's. An interpreter that offers no
+frames reports no place rather than refusing to run.
+
+Diagnostics are safe to log. A shape carries no bound value, so nothing a query
+was asked about appears in a report, and that holds whether or not a parameter
+was marked sensitive.
+
+Diagnostics must not become the thing that exhausts a process, so the number of
+distinct shapes and the number of places per shape are both capped. Executions
+keep being counted after the caps are reached, and shapes seen beyond the cap
+are counted through `unrecorded_shapes`.
+
+An observer is safe to share. A pool hands connections to whichever thread asks,
+so reading a report while several threads record must not fail.
+
+### Query budgets
+
+A repeated query that is only reported is a repeated query that still ships. A
+budget turns the same observation into a refusal:
+
+```python
+from pyoq.diagnostics import QueryBudget, QueryScope
+
+scope = QueryScope(QueryBudget(maximum_repeats=3))
+```
+
+```console
+one statement ran 4 times in this scope, beyond the 3 it was allowed
+from app/teams.py:31 in members: SELECT id FROM person WHERE team_id = ?
+```
+
+`maximum_repeats` is the one that catches an N+1 access, because the shape
+executed once per row of an earlier result is the shape that repeats.
+`maximum_queries` caps the scope as a whole. A refusal names the shape, the
+count, and the place, and carries no bound value. The statement that caused it
+is recorded before it is judged, so a report taken afterwards explains the
+refusal.
+
+Holding a budget costs almost nothing per statement, because only the shape
+just recorded is judged rather than every shape a scope has seen.
+
+### Select-in fetching
+
+Select-in is the strategy that turns an N+1 access into a bounded number of
+statements. `select_in_conditions()` produces one predicate per batch:
+
+```python
+from pyoq.fetching import select_in_conditions
+
+for condition in select_in_conditions(
+    relation, columns, parent_keys, maximum_parameters=999
+):
+    rows = database.many(select(...).from_(source).where(condition))
+```
+
+A key of one column becomes a membership test. A key of several becomes a
+choice between equalities rather than a row value, because every dialect
+renders the first alike and they do not all render the second alike.
+
+The predicate is built from expression nodes rather than through the typed
+query API. Keys come back from a database as values whose types the caller has
+already checked, so routing them through a typed column would fight the column's
+own type without protecting the query the caller wrote.
+
+A contract proves the reduction on SQLite, PostgreSQL, and MySQL alike: the same
+rows come back from far fewer statements than asking once per parent.
+
+### Joined fetching
+
+A to-one relation is one row the join already carries, so it costs no extra
+statement. What it costs is knowing where each table's values land:
+
+```python
+from pyoq.fetching import join_condition, projection_layout
+
+child_at, parent_at = projection_layout((2, 2))
+query = (
+    select(*child_columns, *parent_columns)
+    .from_(table_source(children))
+    .left_join(table_source(parents))
+    .on(join_condition(child_key, parent_key))
+)
+```
+
+Hydration reads a result by position, so the layout a projection is laid out
+with and the layout it is read back by have to be the same one. Deriving both
+from `projection_layout()` keeps them from drifting apart as a plan grows.
+
+A left join leaves a to-one relation absent where nothing matched, which the
+relation states already distinguish from a relation nobody fetched. This is
+proved on SQLite, PostgreSQL, and MySQL alike: one query, hydrated back into
+parents carrying their relation, with the unmatched half absent.
+
+### Nested collections
+
+A nested collection gathers a relation's rows into one value beside their
+parent, so one statement returns the graph:
+
+```console
+SELECT "parent"."id",
+       (SELECT COALESCE(json_group_array(json_object('id', "child"."id")
+                        ORDER BY "child"."id"), json_array())
+        FROM "child" WHERE ("child"."parent_id" = "parent"."id"))
+FROM "parent"
+```
+
+Each dialect gathers rows with its own functions: `json_group_array` and
+`json_object` on SQLite, `jsonb_agg` and `jsonb_build_object` on PostgreSQL,
+`JSON_ARRAYAGG` and `JSON_OBJECT` on MySQL. An empty relation is coalesced to an
+empty collection rather than left null, so a caller need not tell a relation
+with no rows from a fetch that did not happen.
+
+MySQL refuses an ordered collection rather than returning one in the wrong
+order, and says to fetch it with select-in instead. Ordering inside a SQLite
+aggregate needs SQLite 3.44 or later.
+
+A collection reads one table, because a collection is the rows of one relation
+and anything wider is a query rather than a relation.
+
+`decode_collection()` reads the value back into rows. What a driver hands over
+differs: SQLite and MySQL return JSON text while PostgreSQL returns a list it
+has already decoded, and both arrive at the same rows. Values come back in the
+order their names were given, so a caller reads them like any other row.
+
+Nothing is trusted to hold what it promised. A collection that is not a list of
+objects is refused, and a row missing a value it was asked for is reported by
+name and position. A value that is null is a value, because a column that held
+null is not a column that was missing. Values a query did not ask for are left
+alone, so a table gaining a column does not break a fetch.
+
+`hydrate_collection()` turns that into a loaded relation. An empty collection is
+loaded and holds nothing, which is not the same as a relation nobody fetched.
+
+### Relation batching
+
+An N+1 access asks for one parent's children at a time. Gathering the keys first
+turns that into one question, and a dialect's parameter limit turns it into a
+bounded number of questions rather than one per parent:
+
+```python
+from pyoq.relations import RelationBatchLoader
+
+loader = RelationBatchLoader()
+for team in teams:
+    loader.enqueue(members, team.key)
+batches = loader.batches(maximum_parameters=999)
+```
+
+A key spanning several columns costs one bound value per column, so a limit
+counts values rather than keys, and a key wider than the limit is refused rather
+than split. Keys keep their order and never repeat. A key that is null reaches
+nothing, so it is counted through `skipped_keys` rather than asked about, which
+would cost a parameter to learn nothing.
+
+A loader belongs to one scope, and each relation it holds is bounded on its own.
+
+## One chain, from building to reading
+
+A builder describes a statement and never reaches a database. That is what makes
+one safe to share, cache, and test, and it does not change. A context adds the
+other half: it holds the database to run against and hands back builders that
+carry it, so a chain can end in a fetch instead of being handed to an executor.
+
+```python
+from pyoq.dsl import using
+
+dsl = using(database)
+
+dsl.select(TITLE_ID, TITLE_NAME, TITLE_PRICE) \
+   .from_(table_source(TITLES)) \
+   .where(TITLE_PRICE.gt(Decimal("10.00"))) \
+   .order_by(TITLE_ID) \
+   .fetch_all() \
+   .to_list()
+```
+
+A generated table names its own columns, so a query over all of them does not
+have to list them:
+
+```python
+dsl.select_from(TITLES) \
+   .where(TITLE_PRICE.gt(Decimal("10.00"))) \
+   .order_by(TITLE_ID) \
+   .fetch_all() \
+   .into(TitleRow)
+```
+
+The SQL is `SELECT *`. PyOQ still carries the column list, so the row keeps
+its names and each value is read back as the type the column holds, and that
+order is the order the generated row type takes its arguments in. A package
+that no longer matches its schema is refused rather than read by the wrong
+columns.
+
+The shape a caller wants is asked for at the end rather than assembled around
+the call:
+
+```python
+rows = dsl.select(...).from_(...).fetch_all()
+
+rows.to_list()          # [(1, "A Guide", Decimal("12.50")), ...]
+rows.into(Title)        # [Title(1, "A Guide", Decimal("12.50")), ...]
+rows.map(lambda r: r[1])
+rows.to_dicts()         # [{"id": 1, "name": "A Guide", ...}, ...]
+rows.to_dict()          # one row keyed, or an error saying which way it failed
+rows.one()              # exactly one row, or an error saying which way it failed
+rows.first()            # the row at the front, or None
+```
+
+A chain that ends in one row asks the same questions of it. The row is a
+tuple, so it still unpacks, indexes, and compares the way it always did:
+
+```python
+row = dsl.select(TITLE_ID, TITLE_NAME).from_(...).fetch_one()
+
+row.to_dict()           # {"id": 1, "name": "A Guide"}
+row.to_list()           # [1, "A Guide"]
+row.to_tuple()          # (1, "A Guide"), typed as it was selected
+row.into(Title)         # Title(1, "A Guide")
+identifier, name = row  # still a tuple
+```
+
+A row that may be missing is the one shape that had to be tested for before
+it could be built, so the chain takes the type instead:
+
+```python
+dsl.select(...).where(...).fetch_optional_into(Title)   # Title | None
+```
+
+`into` spreads each row across a constructor in the order it was selected, so a
+dataclass or a named tuple needs nothing else. `to_dicts` keys by the name each
+column came back under, and refuses when a column has none rather than inventing
+one.
+
+A query can end in a single row or a single value instead:
+
+```python
+dsl.select(TITLE_NAME).from_(source).where(TITLE_ID.eq(1)).fetch_one()
+dsl.select(TITLE_NAME).from_(source).where(TITLE_ID.eq(1)).fetch_optional()
+dsl.select(TITLE_NAME).from_(source).where(TITLE_ID.eq(1)).fetch_value()
+```
+
+`fetch_value` is only offered to a query that selected one column. That is a
+statement about the type rather than a check when it runs.
+
+Writes chain the same way and end in the number of rows the database wrote:
+
+```python
+dsl.insert_into(TITLES, TITLE_ID, TITLE_NAME).values(1, "A Guide").execute()
+dsl.update(TITLES).set(TITLE_PRICE, Decimal("20.00")).where(TITLE_ID.eq(1)).execute()
+dsl.delete_from(TITLES).where(TITLE_ID.eq(1)).execute()
+
+dsl.insert_into(TITLES, TITLE_ID, TITLE_NAME) \
+   .values(1, "A Guide") \
+   .on_conflict_do_update(TITLE_ID) \
+   .set(TITLE_NAME, "A Guide") \
+   .returning(TITLE_ID, TITLE_NAME) \
+   .fetch_all() \
+   .into(Title)
+```
+
+A set operation orders by a column it selected, because its operands may read
+different tables and a table-qualified name means nothing once they are
+combined. Naming anything else there is refused rather than sent to a server
+that would reject it.
+
+Types are carried the whole way. `select(TITLE_ID, TITLE_NAME, TITLE_PRICE)`
+gives a chain over `tuple[int, str, Decimal]`, `fetch_all().to_list()` is a
+`list[tuple[int, str, Decimal]]`, and `fetch_one()` is the tuple itself.
+
+Unwrapping a bound query with `query`, or a bound write with `statement`, gives
+back the ordinary builder, which cannot reach a database. Anything built that
+way still runs through `dsl.fetch(...)` or `dsl.execute(...)`.
+
+### An asynchronous database
+
+`using` answers the same question either way, so an asynchronous database gives
+an asynchronous context. Only the ending changes:
+
+```python
+rows = await (
+    using(database)
+    .select(TITLE_ID, TITLE_NAME)
+    .from_(table_source(TITLES))
+    .order_by(TITLE_ID)
+    .fetch_all()
+)
+rows.into(Title)
+
+await using(database).insert_into(TITLES, TITLE_ID).values(1).execute()
+```
+
+The chain is built the same way and describes the same statement. `fetch_all`,
+`fetch_one`, `fetch_optional`, `fetch_value`, and `execute` are awaited because
+the database is, and the types carry through exactly as they do synchronously.
+
+## Django
+
+Django 4.2 through 6.x, on any Python each of them supports.
+
+[`docs/django.md`](docs/django.md) is a step by step guide from an empty project
+to typed reads, writes, transactions, policies, and events. What follows is the
+reference.
+
+PyOQ runs on the connection Django already has open:
+
+```python
+from pyoq.django import DjangoOperations
+
+database = DjangoOperations()  # or DjangoOperations("reporting")
+rows = database.many(select(...).from_(source))
+```
+
+```console
+python -m pip install "pyoq-sql[django]"
+```
+
+No pool is opened beside Django's. Two pools against one database is two views
+of what has been committed, and the point of running inside Django is that there
+is one. The connection is resolved for each statement rather than held, because
+Django gives each thread a different one and closes them between requests.
+
+The dialect comes from the connection rather than from the caller, since it is
+the connection that is on the socket. Values are adapted the same way they are
+on PyOQ's own connections, because a backend does not care whose connection it
+is: a decimal or a date still has to reach it in the shape its driver accepts.
+A backend PyOQ has no dialect for is refused by name.
+
+A timeout is applied where the backend has a setting for one, and the session is
+left exactly as it was found. Inside a transaction the setting is local, so the
+database reverts it when the transaction ends; outside one, the previous value is
+read first and put back after. A restore never replaces the failure that made it
+necessary, because a statement that failed can leave a connection unable to
+accept the next one. SQLite has no such setting, and MySQL's covers reads only,
+so a caller is given what the backend can honour rather than a promise it
+cannot keep. A caller's cancellation is honoured before and after a statement
+regardless.
+
+Routing is a question about a model, because that is the only thing a Django
+router is given to decide on:
+
+```python
+database = DjangoOperations.for_model(Report)  # wherever this is routed
+database = DjangoOperations.for_model(Report, write=True)
+```
+
+A caller without a model names the alias instead.
+
+### Transactions
+
+PyOQ takes part in the transaction Django opened and never opens one of its own.
+There is no commit or rollback on `DjangoOperations`, because the block that
+starts a transaction is the one that ends it:
+
+```python
+with database.atomic():
+    database.execute(insert_into(REPORTS, name).values("quarterly"))
+    database.on_commit(lambda: notify("saved"))
+```
+
+Ask the database for its own block rather than reaching for Django's. Django's
+`transaction.atomic()` covers the default alias unless told which to use, so a
+caller working against another database inside a bare block has no transaction
+there at all and nothing reports it. `database.atomic()` always means this
+database. The same holds for `database.on_commit()`, which waits for this
+database to commit.
+
+Nesting a block is a savepoint, and `atomic(savepoint=False)`,
+`atomic(durable=True)`, and `on_commit(robust=True)` behave as Django defines
+them. `database.in_transaction` reports whether a transaction is open.
+
+Test isolation needs nothing extra. A Django test case wraps its test in a block
+and rolls it back, and PyOQ writes through the connection that block belongs to.
+
+An async view reaches a statement through Django's own bridging:
+
+```python
+await sync_to_async(read_reports, thread_sensitive=True)(database)
+```
+
+Django refuses a synchronous statement called directly from a coroutine, and
+PyOQ inherits that guard. Nothing here shares transaction state across the
+boundary, because a transaction belongs to the connection that a thread holds.
+
+### Generating from migrations
+
+Add PyOQ to the project and tell it where generated code belongs:
+
+```python
+INSTALLED_APPS = ["pyoq.django", ...]
+
+PYOQ = {
+    "codegen_directory": "app/_generated",
+    "codegen_package": "app._generated",
+}
+```
+
+```bash
+python manage.py pyoq_codegen
+python manage.py pyoq_codegen catalog --database reporting
+python manage.py pyoq_codegen --dry-run
+python manage.py pyoq_codegen --check
+```
+
+The schema comes from the migration files, not from a database. Migrations are
+what the schema is going to be, and a developer generating types has usually
+not applied them yet. Nothing connects: Django answers what column a field
+takes on a backend without consulting a server, so the alias selects the
+dialect rather than opening a socket.
+
+What that yields is the schema the database will really have, which is not
+always the one the models appear to describe:
+
+- A Django default is applied in Python, so the column has no default. PyOQ
+  says so rather than inviting an insert the database would refuse. The model's
+  default is kept as column metadata.
+- `on_delete` is carried out by Django, and the constraint it writes names no
+  action at all. The relation records what the database will do, with the
+  model's choice beside it.
+- A unique constraint carrying a condition is not a key. It is unique among the
+  rows it matches and not among the others, and treating it as a key would
+  describe a relation reaching many rows as reaching one.
+- A many-to-many field owns a table that migration state never lists, so PyOQ
+  describes it from the field. A through model the project wrote is already a
+  model and is left alone.
+
+Generation itself is the same pipeline every other entry point uses, so
+staging, validation, the manifest, drift detection, and the atomic replacement
+behave identically here.
+
+To generate after every migration, put PyOQ ahead of the app that supplies the
+original command and ask for it:
+
+```python
+PYOQ = {..., "codegen_after_makemigrations": True}
+```
+
+Django's own command runs first and unchanged. A run that writes no migration,
+such as `--dry-run` or `--check`, generates nothing: the files on disk still
+describe the schema before the change, and generating from them would look like
+it had worked.
+
+### A project to read
+
+[`examples/django_project`](examples/django_project) is a project laid out the
+way a real one is: settings, a URLconf, WSGI and ASGI entry points, the admin,
+two applications with models and migrations, a template, and views. It has two
+databases and a router, and its models are chosen to show what reaches the
+database and what stays in Python.
+
+[`examples/django_integration.py`](examples/django_integration.py) applies its
+migrations, generates from them, writes through the ORM and PyOQ on the same
+connection, and then exercises the project through its own views, including an
+async one, along with transactions, savepoints, and commit hooks:
+
+```bash
+python examples/django_integration.py
+```
+
+## FastAPI
+
+The pool belongs to the application, so the lifespan opens it once and gives it
+back once:
+
+```python
+from pyoq.fastapi import SyncDatabase, database_lifespan
+
+def open_notes() -> SyncDatabase:
+    pool = SQLitePool(SQLiteConnectionFactory(path))
+    return SyncDatabase(SQLiteExecutor(pool), pool.close)
+
+app = FastAPI(lifespan=database_lifespan(synchronous={"default": open_notes}))
+```
+
+A database that fails to open takes the application down, and the ones already
+open are given back first. Starting up halfway and serving requests against a
+partly opened application is worse than not starting.
+
+Routes ask for what they need:
+
+```python
+import pyoq.fastapi as pyoq_fastapi
+
+Database = Annotated[QueryOperations, Depends(pyoq_fastapi.database())]
+Transaction = Annotated[QueryOperations, Depends(pyoq_fastapi.transaction())]
+
+@app.get("/notes")
+def list_notes(database: Database) -> dict[str, object]:
+    return {"notes": using(database).select(...).fetch_all().to_dicts()}
+
+@app.post("/notes")
+def add_note(database: Transaction) -> dict[str, int]:
+    return {"written": using(database).insert_into(...).values(...).execute()}
+```
+
+A request that returns commits. A request that raises rolls back, and so does
+one the client gave up on: FastAPI tears the dependency down either way, and the
+transaction is told what tore it down. An `HTTPException` counts as raising,
+because a refused request is still a refused request.
+
+Asking twice for the same dependency gives back the same object, so an override
+matches:
+
+```python
+app.dependency_overrides[pyoq_fastapi.database()] = lambda: substitute
+```
+
+### Which database a route may use
+
+`database()` and `transaction()` are synchronous. `async_database()` and
+`async_transaction()` are their counterparts.
+
+A synchronous transaction belongs to a synchronous route. FastAPI runs a
+synchronous dependency in a worker thread and an async route body on the event
+loop, and a transaction belongs to the thread that opened it. PyOQ refuses the
+mismatch rather than running a statement somewhere it does not belong, so an
+async route takes `async_transaction()`.
+
+### Request budgets
+
+A budget is per request, so one request cannot spend another's allowance:
+
+```python
+Budgeted = Annotated[
+    QueryOperations,
+    Depends(pyoq_fastapi.scoped(QueryBudget(maximum_queries=20))),
+]
+```
+
+Every read and write is recorded before the driver sees it, so the statement
+that goes beyond the allowance never reaches the database. `maximum_repeats` is
+the one that catches a query running once per row of an earlier result.
+
+`ScopedOperations` and `AsyncScopedOperations` do the counting and are not
+specific to FastAPI. Either wraps any database:
+
+```python
+from pyoq.diagnostics import QueryScope, ScopedOperations
+
+scope = QueryScope(QueryBudget(maximum_queries=20))
+counted = ScopedOperations(database, scope)
+```
+
+See [`examples/fastapi_application.py`](examples/fastapi_application.py) for a
+runnable application covering all of it.
+
+## Sanic
+
+Sanic starts its workers as separate processes. A connection made before they
+exist would be shared by all of them, and two processes talking over one socket
+corrupt each other's results. So every pool is opened from the listener that
+runs inside each worker:
+
+```python
+from pyoq.sanic import AsyncDatabase, attach_databases
+
+async def open_notes() -> AsyncDatabase:
+    pool = await PostgresAsyncPool(PostgresAsyncConnectionFactory(dsn)).open()
+    return AsyncDatabase(PostgresAsyncExecutor(pool), pool.close)
+
+attach_databases(app, asynchronous={"default": open_notes})
+```
+
+A database that fails to open takes the worker down, and the ones already open
+are given back first. Every one a worker opened is closed when it stops.
+
+### One transaction per request
+
+A transaction belongs to the block the handler enters:
+
+```python
+import pyoq.sanic as pyoq_sanic
+
+@app.post("/notes")
+async def add_note(request):
+    async with pyoq_sanic.transaction(request) as database:
+        ...
+```
+
+Leaving the block normally commits. Leaving it any other way rolls back: a
+handler that raised, and a request the server abandoned.
+
+It is a block rather than a pair of middlewares for a reason. A server that is
+told the client has gone cancels the task it called the application on, and
+nothing after the handler runs, response middleware included. A transaction
+ended there would keep its connection for as long as the worker lives, inside an
+open transaction, holding whatever it locked. Python ends a block whatever
+happens to the task running it, so the connection goes back to the pool.
+
+A commit that fails raises on the way out of the block, which Sanic answers the
+way it answers any other failure in a handler.
+
+`synchronous_transaction` is the same for a database that is not awaited.
+
+### Request budgets
+
+```python
+pyoq_sanic.attach_request_budget(app, QueryBudget(maximum_queries=20))
+
+@app.get("/notes")
+async def list_notes(request):
+    database = pyoq_sanic.budgeted_of(request)
+```
+
+A budget belongs to the request, so one request cannot spend another's.
+`attach_synchronous_request_budget` and `synchronous_budgeted_of` are the
+counterparts.
+
+See [`examples/sanic_application.py`](examples/sanic_application.py) for a
+runnable application covering all of it.
+
+## Policies
+
+A rule written against SQL text can be walked around with an alias, a subquery,
+or a common table. A policy is written against the statement's own nodes, before
+any SQL exists, so every one of those routes carries it:
+
+```python
+from pyoq.policies import SoftDelete, TenantScope, governed
+
+scoped = governed(database, [TenantScope("tenant_id", 7), SoftDelete("deleted_at")])
+```
+
+```sql
+SELECT COUNT(*) FROM "notes" AS "n"
+WHERE (("n"."tenant_id" = ?) AND ("n"."deleted_at" IS NULL))
+```
+
+The condition follows the name the statement gave the table, joins are scoped
+beside it, a subquery and a common table are scoped inside themselves, and both
+halves of a set operation are scoped. A write carries the scope as a value, so a
+row cannot be created outside the scope that would then be unable to see it, and
+a caller that names the column itself has its value replaced rather than
+honoured. An update or a delete that said it meant every row now means every row
+the policy admits.
+
+### What is available
+
+- `TenantScope(column, value)` puts every read and every write inside one scope
+- `SoftDelete(column)` keeps a row marked deleted out of every read
+- `AllowedTables(names)` refuses a statement that touches anything else
+- `RowConstraint(table, build)` narrows one table by a condition of your own
+
+### What cannot be governed
+
+SQL that was compiled elsewhere has no statement left to read, so it is refused
+rather than quietly allowed. So is raw SQL written into an expression. A project
+that means to run them says so once:
+
+```python
+governed(database, [...], raw_sql=True)
+```
+
+### Transactions and streams
+
+A governed database opens its own transactions and streams its own reads, both
+held to the same rules:
+
+```python
+with scoped.transaction() as active:
+    ...
+with scoped.stream(query) as rows:
+    ...
+```
+
+Without those a caller would reach past the policy to the database underneath to
+open one, and everything inside it would be ungoverned. A dialect's own
+streaming settings belong to the database underneath, which is where it was
+given them.
+
+### Going around a policy
+
+```python
+with administrative_bypass(scoped, "restoring a withdrawn note", audit) as free:
+    ...
+```
+
+The ungoverned database is yielded rather than returned, so it belongs to the
+block and cannot be kept past it. A reason is required, and entering and leaving
+are both written to the audit.
+
+See [`examples/policies.py`](examples/policies.py) for a runnable walkthrough.
+
+## Reading a column as the type it was declared to hold
+
+A driver answers with what its own protocol carries, which is not always what a
+column was declared to be. SQLite hands back a float for a decimal and a string
+for a date, and psycopg hands back a view of its buffer for bytes. A descriptor
+that says `Decimal` and yields a float is a promise the runtime does not keep,
+and money is the usual casualty.
+
+A descriptor carries the type it was declared with, so a row is given back as
+the row was described:
+
+```python
+PRICE = ColumnDescriptor[Decimal](..., value_type=Decimal)
+
+row = using(database).select(PRICE).from_(source).fetch_one()
+# Decimal('12.50'), not 12.5
+```
+
+Generated descriptors declare it for you. `Decimal`, `date`, `datetime`,
+`time`, `timedelta`, `UUID`, `bytes`, `bool`, `int`, `float`, `str`, and JSON
+are read from whatever a driver answered with. A generated enumeration reads
+back as a member of itself rather than as the string stored for it, and an
+array column reads back as the tuple its descriptor declares rather than as
+the list a driver hands over.
+
+This happens on every path that answers with rows: `fetch_one`, `fetch_all`,
+scalars, bulk statements that return rows, streams, and the asynchronous
+counterpart of each. A driver decides what it carries and decoding decides
+what a caller is given, so two paths that decode differently would answer the
+same query with different types.
+
+A column that declared nothing is handed back exactly as it came, so a caller
+that never said what it wanted pays nothing for the question.
+
+A value that cannot be read as what it was declared to be raises, because
+handing back the wrong type silently is the defect this exists to stop. What a
+driver already lost is lost: PyOQ converts what it is given and cannot recover
+precision a database did not keep.
+
+A nested collection arrives as JSON, which spells a date as a string and a
+decimal as a number, so it is told what its columns hold in the same way:
+
+```python
+decode_collection(payload, ("id", "released"), (int, date))
+hydrate_collection(payload, ("id", "released"), construct, (int, date))
+```
+
+## Observability
+
+Instrumentation is a database that wraps another one, so a project that wants
+none holds the plain database and pays nothing at all:
+
+```python
+from pyoq.diagnostics import CollectingSink, InstrumentedOperations
+
+watched = InstrumentedOperations(database, sink)
+```
+
+Every statement is reported before it reaches a driver and again when it is
+answered, at the one place they all pass. An event carries the shape of the
+statement, how many parameters it had, how long it took, and how many rows came
+back.
+
+### What an event may carry
+
+Nothing a query was asked about, unless a policy says so:
+
+```python
+InstrumentedOperations(database, sink, EventPolicy(
+    include_values=True,          # never the ones marked sensitive
+    include_failure_detail=True,  # a driver names the offending value in it
+    slow_after=0.5,
+))
+```
+
+A failure is reported by the name of its type. PostgreSQL states the value that
+caused it inside the message it raises, so the message is withheld until a
+project decides otherwise. A value the compiler marked sensitive is left out
+even when a policy asks for values, because marking it was the decision that it
+must not be shown.
+
+A shape has its literals taken out as well as its placeholders, so a statement
+written by hand can be logged as safely as one PyOQ built. The statement is read
+once, left to right, because what a character means depends on what it is
+inside: a quote inside a comment starts nothing, and two dashes inside a string
+are not a comment. Quoted strings, dollar quoted strings, and every kind of
+comment go, nested ones whole; quoted names stay, because a name is not a value.
+A number is taken out however it was written, and an opening that was never
+closed takes the rest of the statement with it, because nothing after it can be
+shown to be safe. What a shape carries is
+bounded, while its digest is taken from the whole statement, so two that differ
+only past the bound are still told apart.
+
+### Tracing
+
+```python
+from pyoq.tracing import TracingSink
+
+watched = InstrumentedOperations(database, TracingSink())
+```
+
+Each answered statement becomes a span carrying the same facts the event does.
+A span is opened and closed when the statement is answered, because that is the
+event that knows how long it took.
+
+### Readings
+
+```python
+from pyoq.diagnostics import metrics_of, shape_cache_metrics
+
+metrics_of(pool)          # open, idle, checked out, and how much is in use
+shape_cache_metrics()     # how often a shape was already known
+```
+
+Events say what happened; a reading says what is happening now. The shape cache
+is bounded, because the statements a process runs are not.
+
+See [`examples/observability.py`](examples/observability.py) for a runnable
+walkthrough, and `benchmarks/observability.py` for the cost it is held to.
+
+## Generated names
+
+`NamingPolicy` converts source identifiers into valid Python names as one
+immutable batch. It supports snake case, Pascal case, and upper snake case.
+Normalization uses Unicode NFKC and Python's locale-independent case rules.
+Python keywords, soft keywords, and names reserved by a generated component
+receive trailing underscores until the name is available.
+
+Each request has a stable key chosen by the generation planner. Scopes isolate
+namespaces such as table types, row fields, and enum members:
+
+```python
+from pyoq.naming import (
+    NameRequest,
+    NamingPolicy,
+    NamingScope,
+    PythonNameStyle,
+)
+from pyoq.schema import Identifier
+
+fields = NamingScope(
+    name="user-fields",
+    style=PythonNameStyle.SNAKE_CASE,
+    requests=(
+        NameRequest("first-name", Identifier("First Name")),
+        NameRequest("class", Identifier("class")),
+    ),
+    reserved_names=("builder", "build"),
+)
+names = NamingPolicy().resolve((fields,))
+
+assert names.get("user-fields", "first-name") == "first_name"
+assert names.get("user-fields", "class") == "class_"
+```
+
+Ordering requests or scopes differently does not change the result. Names that
+become equal after case conversion, Unicode normalization, punctuation removal,
+or reserved-name escaping are not renamed by sequence. Instead,
+`NamingCollisionError` reports every conflict in the batch before a renderer
+can receive a partial result. This keeps generated APIs stable when database
+reflection order changes.
+
+## Schema snapshots and drift
+
+A snapshot is the one canonical record of what a database holds. Written to a
+file it becomes reviewable in a diff and usable without a database, which is
+what lets generation and staleness checks run where no credentials exist.
+
+```toml
+[tool.pyoq]
+schema-snapshot = "schema.json"
+```
+
+With that set, generation reads the file instead of connecting. Nothing else
+changes: the same pipeline serves both and does not know which it got.
+
+```bash
+pyoq snapshot   # record what the database holds
+pyoq drift      # compare the record against the database
+pyoq generate   # generate from the record, with no connection
+```
+
+Recording the same database twice writes the same bytes, so the file changes
+only when the schema does. It is written beside its destination and moved into
+place, so a reader never sees half a snapshot and a failed write leaves the
+previous record intact.
+
+`pyoq drift` names what moved in terms of the schema rather than of the file,
+so the answer is the column that changed rather than the line that differs.
+
+```text
+1 schema difference(s): added column warehouse.public.parts.weight
+```
+
+Named things are matched by name and then by value, so a renamed table reads
+as one removed and one added rather than as a change to something that is no
+longer there.
+
+### After a migration
+
+A migration changes the schema, which makes the generated package stale and
+the recorded snapshot wrong. One call puts both back in step, in that order,
+because a package regenerated before the snapshot is written would be built
+from the schema that has just been replaced.
+
+```python
+from pyoq.migrations import after_migration_at
+
+after_migration_at("/path/to/project")
+```
+
+Nothing there imports a migration tool, and no migration tool is a dependency
+of PyOQ.
+
+It runs after the migration has been applied, which for Alembic means after
+`alembic upgrade`. A post-write hook is not that moment: Alembic runs those
+inside `alembic revision`, on the revision file it has just written, while
+the database is still whatever it was. Wiring PyOQ there would record the
+schema the migration is about to replace, so the installed command refuses a
+revision file and says where the step belongs.
+
+Call it from `env.py`, once the migrations have run:
+
+```python
+with connectable.connect() as connection:
+    context.configure(connection=connection, target_metadata=target_metadata)
+    with context.begin_transaction():
+        context.run_migrations()
+after_migration_at(PROJECT_ROOT)
+```
+
+Or run it as a step of its own:
+
+```bash
+alembic upgrade head && pyoq-after-migration
+```
+
+The command takes the project directory, or nothing and uses the working
+directory, walking upwards to the first directory holding a `pyproject.toml`.
+Django, a shell script, or a CI job call `after_migration_at` directly and get
+the same two steps.
+
+## Generation pipeline
+
+Generation is a staged application service assembled from small typed
+components. Schema loading, concern rendering, syntax validation, drift
+inspection, locking, cleanup, manifest storage, and package replacement have
+independent interfaces and one responsibility each.
+
+```mermaid
+flowchart LR
+    Source[Schema source] --> Snapshot[Immutable snapshot]
+    Snapshot --> Planner[Generation planner]
+    Planner --> Renderers[Concern renderers]
+    Renderers --> Plan[Canonical plan]
+    Plan --> Validator[Python syntax validator]
+    Validator --> Lock[Project lock]
+    Lock --> Drift[Drift checker]
+    Drift --> DryRun[Dry run report]
+    Drift --> Check[Drift check]
+    Drift --> Writer[Atomic writer]
+    Writer --> Stage[Staged package]
+    Stage --> Manifest[Ownership manifest]
+    Manifest --> Swap[Package replacement]
+```
+
+Every rendered path is relative and portable. Plans sort paths before
+validation, normalize line endings, reject duplicate outputs from any concern,
+and hash UTF-8 bytes with SHA-256. Python and stub files are parsed before the
+project lock or generated directory can be changed.
+
+The ownership manifest is compact, versioned, and deterministic. A generated
+file is replaceable or removable only when its path is recorded and its current
+checksum still matches the manifest. Existing paths without ownership records,
+modified generated files, symlinks, and unsafe parent paths fail closed.
+Unowned files elsewhere in the generated directory are copied forward without
+content changes. Cleanup considers only stale manifest entries.
+
+Generation uses a fail-closed project lock at `.pyoq-generation-lock`. If a
+process is interrupted without releasing it, verify that no generation command
+is active before removing that directory.
+
+Write mode copies the current package into a sibling staging directory, applies
+the validated plan there, and replaces the destination package only after the
+new manifest is complete. A failed replacement restores the previous package.
+Check mode fails on creates, updates, removals, ownership conflicts, or manifest
+drift. Dry-run mode returns the same categorized change counts without writing
+the generated package.
+
+The complete in-memory example assembles every production component and runs
+dry-run, write, and check modes without a database or network connection:
+
+```console
+python examples/generation_pipeline.py
+```
+
+See [`examples/generation_pipeline.py`](examples/generation_pipeline.py) for
+the typed schema source and pipeline assembly. Database reflection is
+introduced in a later package phase. Until a built-in schema source is
+available, the default command-line services continue to report generation as
+unavailable.
+
+## Generated database types
+
+`GeneratedTypesRenderer` turns one immutable schema snapshot into a complete
+Python package. The renderer resolves all names and type mappings once, then
+passes that canonical model to focused file renderers.
+
+```mermaid
+flowchart LR
+    Snapshot[Schema snapshot] --> Model[Canonical generation model]
+    Model --> Enums[enums.py]
+    Model --> Rows[rows.py]
+    Model --> Writes[writes.py]
+    Model --> Tables[tables.py]
+    Model --> Relations[relations.py]
+    Model --> Facade[package facade]
+```
+
+Add the renderer as one concern in the generation planner:
+
+```python
+from pyoq.generation import GeneratedTypesRenderer, GenerationPlanner
+
+planner = GenerationPlanner((GeneratedTypesRenderer(),))
+```
+
+For a `user` table, generated names follow these roles:
+
+| Generated name | Responsibility |
+|---|---|
+| `User` and `USER` | Typed table descriptor and its shared instance |
+| `UserRow` | Frozen result value containing every readable column |
+| `UserInsert` | Frozen insert value with required and omitted-field semantics |
+| `UserUpdate` | Frozen update value where every writable field may be omitted |
+| `UserInsertValues` | Required and optional dictionary shape for typed boundaries |
+| `UserUpdateValues` | Optional dictionary shape for update boundaries |
+| `UserPrimaryKey` | Frozen primary-key value |
+| `UserBuilder` | Immutable insert builder without `build()` until complete |
+| `UserUpdateBuilder` | Immutable update builder with concrete field setters |
+
+Table and column descriptors support SQL-shaped discovery. Constants use the
+source database name while carrying its exact value as metadata:
+
+```python
+from application.database import USER, User
+
+identifier_column = USER.ID
+new_values = User.builder().tenant_id(7).name("Ada").build()
+```
+
+Every writable column produces a concrete setter with its exact mapped Python
+type. Nullable setters accept `None`. Generated, identity, computed, and other
+read-only columns expose no insert or update setter. Calling a setter returns a
+new frozen builder and never performs database I/O.
+
+Required fields use `Missing` and `Present` type states. Each required setter
+adds two overloads, so generated typing grows linearly with required-column
+count. The final required setter returns a distinct complete-builder subtype.
+Only that subtype defines `build()`. Mypy and Pyright therefore reject both an
+empty build and a partially complete build, while editors can omit `build()`
+from their completion lists. Runtime incomplete builder objects also lack that
+attribute.
+
+`UserInsert` is a reusable value object. The column-oriented insert statement
+API accepts it through `values_many()` alongside positional multi-row values.
+Constructing it has no connection, transaction, or persistence side effect.
+
+Each generated column descriptor records both its database name and the Python
+field name used by the generated row and write values. That pairing is what
+lets a write statement map a generated value onto the columns it selected
+without dynamic name guessing.
+
+Primary and unique keys become frozen exact-type values. Relationships become
+typed descriptors containing aligned source and target columns, referential
+actions, and source names. Relationship descriptors describe metadata only;
+they do not trigger lazy loading or hidden queries.
+
+SQL scalars map to narrow Python types, including `Decimal`, `UUID`, date and
+time values, immutable tuples for arrays, generated `StrEnum` classes, and a
+recursive `JsonValue` alias. Unknown source types map to `object`, never
+`Any`. Generated source is deterministic and passes Ruff formatting, Mypy
+strict mode, and Pyright strict mode without suppressions.
+
+A domain is a named type with rules attached, and it maps to whatever it is
+written over: a domain over `text` generates `str`, and one over
+`numeric(12, 2)` generates `Decimal`. The schema keeps the name as well as the
+base, because a schema that forgot it would no longer describe the database it
+was read from. A domain that forbids null makes every column of it not
+nullable, whatever the column itself said, because the server refuses a null
+there either way. PostgreSQL has domains; SQLite and MySQL have none.
+
+Run the in-memory rendering example without a database or network connection:
+
+```console
+python examples/generated_types.py
+```
+
+## Typed expressions
+
+Fields, bound values, computed expressions, and conditions form an immutable
+typed expression tree. Named methods keep SQL semantics explicit and preserve
+the result type in Mypy and Pyright strict modes.
+
+```python
+from decimal import Decimal
+
+from pyoq.query import bind, field
+
+USER_ID = field(int, "id", table_name="users")
+USER_NAME = field(str, "name", table_name="users")
+USER_BALANCE = field(Decimal, "balance", table_name="users")
+USER_TENANT = field(int, "tenant_id", table_name="users")
+
+predicate = USER_ID.gt(0) & USER_NAME.starts_with("A")
+adjusted = USER_BALANCE.add(Decimal("5.00"))
+selected = USER_ID.in_(bind(1), bind(2))
+```
+
+Generated column descriptors implement the same `Expression[T]` contract, so
+the generated database package is the primary field source. Comparisons,
+numeric arithmetic, string operations, temporal extraction and duration
+arithmetic, null predicates, ranges, membership, and boolean composition are
+available without converting descriptors or losing their value types.
+
+```mermaid
+flowchart LR
+    Field[Typed field] --> Node[Immutable expression node]
+    Value[Python value] --> Bind[Bound value node]
+    Bind --> Node
+    Node --> Computed[Typed computed expression]
+    Node --> Condition[Boolean condition]
+    Raw[Explicit typed raw template] --> Node
+```
+
+Python values supplied to expression methods always become bound-value nodes.
+They cannot become field names, operators, clauses, or raw SQL text. Null
+equality is normalized to `is_null()` or `is_not_null()` nodes. Explicit raw
+expressions use named placeholders that accept expression objects only:
+
+```python
+from pyoq.query import raw
+
+distance = raw(
+    float,
+    "distance({origin}, {target})",
+    origin=USER_ID,
+    target=bind(10),
+)
+```
+
+Raw placeholders reject plain values, attribute access, indexing, conversion,
+and format specifications. Use `raw_condition()` when the template produces a
+boolean condition. SQL rendering and ordered parameter extraction are owned by
+the compiler introduced with statement construction.
+
+Run the expression example with:
+
+```console
+python examples/expressions.py
+```
+
+### Choosing one value out of several
+
+`case()` reads its branches in order and stops at the first that holds. The
+type of the whole expression is the type of its first result, so a later branch
+that disagrees is refused where it is written.
+
+```python
+from pyoq.query import case, coalesce, greatest, least, nullif
+
+tier = (
+    case()
+    .when(USER_BALANCE.gt(Decimal("500.00")), "premium")
+    .when(USER_BALANCE.gt(Decimal("100.00")), "standard")
+    .otherwise("basic")
+)
+```
+
+`otherwise()` gives the value for rows no branch claimed, and the result is not
+nullable. `end()` closes the case without one, and the result is `T | None`,
+because SQL answers null for a row nothing matched:
+
+```python
+flagged = case().when(USER_BALANCE.lt(Decimal("0.00")), "overdrawn").end()
+```
+
+The other three answer the same question with fixed rules. `coalesce()` takes
+the first argument that is not null, and it drops the `None` from the type when
+the fallback cannot be null. `USER.NICKNAME` below is a generated descriptor for
+a nullable column, typed `Expression[str | None]`:
+
+```python
+from generated.database.tables import USER
+
+label = coalesce(USER.NICKNAME, "unknown")
+```
+
+`label` is `ComputedExpression[str]`, so a row read through it needs no null
+check and no cast. `nullif()` answers null when its two arguments agree, `greatest()` takes the
+widest of its arguments and `least()` the narrowest:
+
+```python
+blank_as_null = nullif(USER_NAME, "")
+floor_price = greatest(USER_BALANCE, Decimal("0.00"))
+capped = least(USER_BALANCE, Decimal("1000.00"))
+```
+
+Every one of these is an expression like any other, so it can be selected,
+ordered by, grouped by, compared, or nested inside another. Values reach them
+as bound parameters, never as SQL text. SQLite has no `GREATEST` or `LEAST` and
+spells them `MAX` and `MIN` over several arguments, which the SQLite compiler
+emits without changing what the expression means.
+
+### Asking for a value as another type
+
+`cast()` gives the conversion to the database and declares what it answers
+with, so the value is read back as that type rather than as whatever the driver
+carried:
+
+```python
+from pyoq.query import cast
+
+as_number = cast(USER_NAME, Decimal)
+as_text = cast(USER_ID, str)
+```
+
+`as_number` is `ComputedExpression[Decimal]`, and a row read through it holds a
+`Decimal` on every dialect. A cast is the only computed expression that names
+its own type, so it is the only one the row decoder can hold to a promise.
+
+Each dialect names its own types, and the names were taken from running
+servers rather than from a specification:
+
+| Target | SQLite | PostgreSQL | MySQL |
+| --- | --- | --- | --- |
+| `bool` | `INTEGER` | `BOOLEAN` | refused |
+| `int` | `INTEGER` | `INTEGER` | `SIGNED` |
+| `float` | `REAL` | `DOUBLE PRECISION` | `DOUBLE` |
+| `Decimal` | `NUMERIC` | `NUMERIC` | `DECIMAL(65, 30)` |
+| `str` | `TEXT` | `TEXT` | `CHAR` |
+| `bytes` | `BLOB` | `BYTEA` | `BINARY` |
+| `date` | refused | `DATE` | `DATE` |
+| `time` | refused | `TIME` | `TIME` |
+| `datetime` | refused | `TIMESTAMP` | `DATETIME` |
+| `timedelta` | refused | `INTERVAL` | refused |
+| `UUID` | refused | `UUID` | refused |
+| `dict`, `list` | refused | `JSONB` | `JSON` |
+
+A dialect that has no such type refuses the cast. SQLite accepts
+`CAST(x AS DATE)` and answers with an integer, because an unfamiliar type name
+falls back to storage affinity there instead of being rejected, so PyOQ
+refuses rather than emitting a name that would return the wrong kind of value
+under a promise of the right one. MySQL is asked for its largest decimal
+precision, because a bare `DECIMAL` means `DECIMAL(10, 0)` and would answer
+`123` for `123.45`.
+
+A target no database type answers to is refused where it is written, not when
+the query runs.
+
+### Measuring a row against the rows around it
+
+An aggregate collapses a group into one row. A window leaves the rows alone
+and gives each one an answer computed from its neighbours, so a rank, a
+running total, or the previous row's value can be selected beside the row
+itself.
+
+The clauses read in the order SQL writes them: the function, then the window
+it looks through, then how that window is divided and ordered.
+
+```python
+from pyoq.query import dense_rank, lag, ntile, rank, row_number, sum_
+
+position = row_number().over().partition_by(USER_TENANT).order_by(
+    USER_BALANCE.desc()
+)
+standing = rank().over().order_by(USER_BALANCE.desc())
+running = sum_(USER_BALANCE).over().partition_by(USER_TENANT).order_by(USER_ID)
+previous = lag(USER_BALANCE).over().order_by(USER_ID)
+```
+
+`row_number()` counts from one and breaks ties arbitrarily. `rank()` gives
+tied rows the same position and leaves a gap after them; `dense_rank()` gives
+them the same position and leaves no gap. `ntile(n)` says which of `n` equal
+buckets a row falls in. `lag()` and `lead()` read the value that many rows
+back or ahead, and answer null past the edge of the window, so they are typed
+`T | None`.
+
+Any aggregate can be taken over a window instead of over a group, through the
+same `over()`. `count().over()` counts every row in the window without naming
+a column.
+
+A window with no `partition_by()` covers every row. A window orders rows the
+way a query does, using the same terms and the same rules, so asking for nulls
+last inside a window works on every dialect just as it does outside one.
+
+Window functions require SQLite 3.25 or later. They are governed by the
+`window_functions` capability, so a build without them refuses the query
+rather than emitting SQL it cannot run.
+
+### Comparing several values as one
+
+A composite key is one key, and asking whether a row is among a set of them
+should read that way. `row()` puts columns side by side and compares them
+against tuples in one predicate, instead of an OR of ANDs a reader has to
+reassemble:
+
+```python
+from pyoq.query import row
+
+wanted = row(USER_TENANT, USER_NAME).in_((1, "Ada"), (2, "Grace"))
+exact = row(USER_TENANT, USER_NAME).eq((1, "Ada"))
+after = row(USER_TENANT, USER_ID).gt((1, 100))
+```
+
+The tuples are checked against the columns by position, so a value of the
+wrong type, in the wrong order, or a tuple of the wrong width is refused where
+it is written rather than when the query runs. `row(USER_TENANT,
+USER_NAME).in_((1, 2))` does not type-check, because the second column is a
+string.
+
+`in_()`, `not_in()`, `eq()`, `ne()`, `gt()`, `ge()`, `lt()`, and `le()` are
+available. The ordering comparisons compare left to right, the way SQL orders
+a row value, so `(tenant, id) > (1, 100)` means every row of a later tenant
+and the rows of tenant one after id 100. Matching an empty set of rows is the
+same nothing that an empty `in_()` already means.
+
+A row value is two columns or more. SQL reads a single bracketed value as that
+value, so a row of one is refused. Every value travels bound, exactly as it
+does in any other predicate.
+
+Membership in the rows of another query is a `semi_join()` rather than a row
+value, because that is where the join conditions and their typing already
+live.
+
+### Holding the rows a query read
+
+A row read inside a transaction can be changed by somebody else before the
+transaction acts on it. A lock holds it until the transaction ends, and the
+clauses read in the order SQL writes them:
+
+```python
+job = (
+    select(JOB_ID, JOB_PAYLOAD)
+    .from_(JOBS)
+    .where(JOB_STATE.eq("pending"))
+    .limit(1)
+    .for_update()
+    .skip_locked()
+)
+```
+
+That is a work queue: each worker takes a row nobody else holds, and passes
+over the ones already taken instead of waiting behind them.
+
+`for_update()` holds the whole row. `for_share()` holds it against change
+while letting others read it. `for_no_key_update()` and `for_key_share()` are
+the weaker PostgreSQL locks that leave a key referenceable.
+
+`nowait()` fails instead of waiting for a row somebody else holds, and
+`skip_locked()` passes over it. Waiting is what a lock does when told neither,
+so nothing is written for it. `of()` narrows the lock to some of what the
+query read, which is what keeps a join from holding rows it only looked at.
+
+Each dialect refuses what it has not got, measured against running servers.
+SQLite locks the whole database rather than rows, so it refuses locking
+outright and points at a transaction. MySQL has `FOR UPDATE` and `FOR SHARE`
+with `NOWAIT`, `SKIP LOCKED`, and `OF`, but no weaker lock, so it refuses
+`for_no_key_update()` and `for_key_share()`. PostgreSQL has all four.
+
+A lock belongs to the query that read the rows, so a set operation carries
+none.
+
+### Reading inside a JSON value
+
+A path is steps rather than text, because the three dialects do not agree on
+how a path is written, and PostgreSQL reads another's spelling as a member
+that is simply absent and answers null without complaining. A string step is a
+member and an integer step is an element:
+
+```python
+country = EVENT_BODY.json_text("actor", "country")
+first_tag = EVENT_BODY.json_text("tags", 0)
+tag_count = EVENT_BODY.json_length("tags")
+```
+
+`json_get()` answers with JSON and `json_text()` with what that JSON says, so
+`json_text()` is typed `str | None` and gives you text on every dialect even
+where the driver would have handed back the number it looked like.
+
+`json_has()` asks whether anything is at a path at all, and `json_contains()`
+asks whether one JSON value holds another:
+
+```python
+verified = EVENT_BODY.json_has("actor", "verified")
+from_london = EVENT_BODY.json_contains({"actor": {"city": "London"}})
+```
+
+The path is a value, so it is bound like any other and never enters the SQL.
+Each dialect writes what it has:
+
+| Asked | SQLite | PostgreSQL | MySQL |
+| --- | --- | --- | --- |
+| `json_get` | `-> '$.a.b'` | `#> '{a,b}'` | `JSON_EXTRACT` |
+| `json_text` | `JSON_EXTRACT` | `#>> '{a,b}'` | `JSON_UNQUOTE(JSON_EXTRACT(...))` |
+| `json_length` | `JSON_ARRAY_LENGTH` | `JSONB_ARRAY_LENGTH` | `JSON_LENGTH` |
+| `json_has` | `JSON_TYPE(...) IS NOT NULL` | `#> ... IS NOT NULL` | `JSON_CONTAINS_PATH` |
+| `json_contains` | refused | `@>` | `JSON_CONTAINS` |
+
+SQLite has no containment operator, so it refuses rather than emulating one.
+Asking any of this of a column that is not JSON is refused where it is
+written.
+
+### Asking about an array
+
+An array column is generated as `tuple[T, ...]`, and these questions are typed
+by that element type, so a value of the wrong kind is refused where it is
+written:
+
+```python
+tagged = POST_TAGS.has("python")
+untagged = POST_TAGS.lacks("draft")
+both = POST_TAGS.contains_all(("python", "sql"))
+any_of = POST_TAGS.overlaps(("python", "rust"))
+inside = POST_TAGS.contained_by(("python", "sql", "rust"))
+first = POST_TAGS.element(0)
+count = POST_TAGS.length()
+```
+
+**Elements are counted from zero.** The column reads back as a tuple and a
+tuple counts from zero, so `POST_TAGS.element(0)` is the same element as
+`row.tags[0]`. SQL counts an array from one, and PyOQ writes that difference
+out rather than leaving a caller to remember it. Past the end there is no
+element, so the answer is null and the type is `T | None`.
+
+`length()` is the same method that measures text, because how long a thing is
+is one question whichever kind of thing it is.
+
+Only PostgreSQL has an array type. SQLite and MySQL refuse these and say to
+hold a collection as JSON or in a table of its own, rather than emulating an
+array they have not got.
+
+### Calling a function this database has
+
+Every database grows functions the others have not got, and a toolkit that
+offered only what all three share would be smaller than any of them. A vendor
+function is declared once with the type it answers with, and called like any
+other expression:
+
+```python
+from pyoq.config import DatabaseDialect
+from pyoq.query import vendor_function, vendor_predicate
+
+similarity = vendor_function(
+    float, "similarity", dialects=(DatabaseDialect.POSTGRES,)
+)
+starts_with = vendor_predicate(
+    "starts_with", dialects=(DatabaseDialect.POSTGRES,)
+)
+
+close = similarity(USER_NAME, bind("Ada")).gt(0.3)
+prefixed = starts_with(USER_NAME, bind("Ad"))
+```
+
+`dialects` says which databases the function exists on, and a compiler for any
+other refuses the query rather than sending SQL that cannot run. A declaration
+that names no dialect is written wherever it is asked for, because saying
+nothing means the caller did not say.
+
+`vendor_function()` answers with the type given; `vendor_predicate()` answers
+with a `Condition`, so it composes with `and_()`, `or_()`, and `where()` like
+any other predicate.
+
+Arguments are expressions and travel bound, exactly as they do everywhere
+else. The **name** is not a value, because no database takes a function name
+as a parameter, so it is written into the SQL and held to being a name: it
+must be an identifier, optionally qualified by the schema that holds it.
+Anything else is refused where it is declared.
+
+Use `raw()` instead when what you need is not a call at all but a fragment of
+SQL with a shape of its own.
+
+### Running a stored procedure
+
+A procedure is invoked rather than selected from, so it is a statement rather
+than an expression:
+
+```python
+from pyoq.query import call
+
+database.execute(call("record_one", bind(7)))
+```
+
+Arguments travel bound like any other value. The name is written into the SQL,
+because no database takes a routine name as a parameter, so it is held to
+being an identifier optionally qualified by the schema that holds it.
+
+A procedure can answer with rows, and a schema does not describe what they
+hold, so nothing can be inferred. Naming the columns is what lets the rows be
+read back as the types they were said to be:
+
+```python
+rows = database.many(call("read_one", bind(0)).returning(RECORDED_VALUE))
+```
+
+PostgreSQL has no result set from a procedure and passes a value back through
+an `INOUT` parameter, which `CALL` answers with as one row. MySQL answers with
+whatever the procedure selected. SQLite keeps no stored procedures at all and
+refuses, because there is nothing for it to run and nothing to fall back on.
+
+A stored **function** is a function, not a statement, so it is declared with
+`vendor_function()` and called wherever an expression goes.
+
+## SELECT queries
+
+Every join type is its own method and the clause qualifying it comes after,
+so a chain reads in the order SQL is written:
+
+```python
+select(TITLE_NAME, PUBLISHER_NAME).from_(TITLES).left_join(PUBLISHERS).on(
+    TITLE_PUBLISHER_ID.eq(PUBLISHER_ID)
+)
+```
+
+| SQL | PyOQ |
+| --- | --- |
+| `INNER JOIN t ON c` | `.join(t).on(c)`, or `.inner_join(t).on(c)` |
+| `LEFT JOIN t ON c` | `.left_join(t).on(c)` |
+| `RIGHT JOIN t ON c` | `.right_join(t).on(c)` |
+| `FULL JOIN t ON c` | `.full_join(t).on(c)` |
+| `CROSS JOIN t` | `.cross_join(t)` |
+| `NATURAL JOIN t` | `.natural_join(t)` |
+| `NATURAL LEFT JOIN t` | `.natural_left_join(t)` |
+| `NATURAL RIGHT JOIN t` | `.natural_right_join(t)` |
+| `NATURAL FULL JOIN t` | `.natural_full_join(t)` |
+| `JOIN t USING (a, b)` | `.join(t).using("a", "b")` |
+| `WHERE EXISTS (...)` | `.semi_join(t).on(c)`, or `where(exists(q))` |
+| `WHERE NOT EXISTS (...)` | `.anti_join(t).on(c)`, or `where(not_exists(q))` |
+| `JOIN LATERAL (...)` | `.cross_join(q.as_lateral("name"))` |
+
+A join is qualified once, by `ON`, `USING`, or `NATURAL`. `on_key()` takes a
+generated relationship descriptor and builds the `ON` from the foreign key the
+schema already declares, including one equality per column of a composite key.
+
+A cross join and a natural join finish the chain on their own. Every other
+kind is not a query until it is qualified, which both type checkers enforce.
+
+A semi join keeps rows that have a match without bringing the match back, and
+an anti join keeps rows that have none. No database writes either as a join,
+so neither does the SQL: they become `EXISTS` and `NOT EXISTS`, which every
+supported dialect understands. They neither collide with `where` nor depend
+on being written after it.
+
+A lateral source may read the rows to its left, one row at a time. SQLite has
+none, so it refuses the query rather than evaluating the subquery once and
+quietly meaning something else.
+
+A recursive table declares typed columns before either term is built. The
+column object is reused when reading from the recursive source, so autocomplete
+and static checking cannot replace its type with a caller assertion:
+
+```python
+from pyoq.query import bind, column, recursive_table, select
+
+number = column("n", int)
+walk = recursive_table("walk", number)
+walk_definition = walk.define(
+    select(bind(1)),
+    select(walk.field(number).add(1))
+    .from_(walk)
+    .where(walk.field(number).lt(5)),
+)
+numbers = select(walk.field(number)).with_(walk_definition).from_(walk_definition)
+```
+
+A write returns every column its table declares with `returning_all()`, which
+reads the list generation put on the table.
+
+`select()` retains the exact ordered projection tuple through eight fields.
+Every clause returns a new query and leaves its input reusable. Fields and
+generated column descriptors are the only normal structural inputs, while
+Python values remain bound expression values.
+
+```python
+from pyoq.descriptors import TableDescriptor
+from pyoq.query import count, field, select
+
+USERS = TableDescriptor[object, object, object]("users")
+USER_ID = field(int, "id", table_name="users")
+USER_NAME = field(str, "name", table_name="users")
+
+active_users = (
+    select(USER_ID.as_("user_id"), USER_NAME, count())
+    .from_(USERS)
+    .where(USER_ID.gt(0))
+    .group_by(USER_ID, USER_NAME)
+    .having(count().gt(0))
+    .order_by(USER_NAME.asc())
+    .limit(20)
+)
+```
+
+A query source is a table descriptor instance, never a generated table class.
+The generated class is the typed column and builder namespace, and the
+generated constant beside it is the value that `from_()`, `join()`,
+`cross_join()`, and `table_source()` accept. Passing the class is rejected by
+both strict type checkers and, for untyped callers, by a `QueryValidationError`
+that states the instance requirement.
+
+The immutable model covers projection aliases, distinct selection, table and
+aliased-table sources, inner and outer joins, cross joins, predicates,
+grouping, aggregate filters, ordering with explicit null placement,
+pagination, derived-table subqueries, common table expressions, and set
+operations. Aggregates preserve numeric and scalar result types while marking
+empty-set results nullable where required.
+
+```mermaid
+flowchart LR
+    Projection[Typed projections] --> Select[Immutable SELECT node]
+    Source[Table or derived source] --> Select
+    Join[Typed joins] --> Select
+    Predicate[Boolean conditions] --> Select
+    Select --> Derived[Subquery or common table]
+    Select --> Set[Typed set operation]
+```
+
+Set operations require the same projection type and validate projection counts
+again at runtime. Common table column lists must match the selected arity.
+Structural identifiers reject empty strings and null characters. The model
+does not render or execute SQL; compilation and execution own those separate
+responsibilities.
+
+Run the SELECT construction example with:
+
+```console
+python examples/select_queries.py
+```
+
+## Typed INSERT statements
+
+`insert_into(table, *columns)` mirrors SQL and types the values by the selected
+columns. Each `values()` call adds one row, and repeated calls compile into a
+single multi-row statement rather than separate statements. Statement
+construction is immutable and performs no I/O.
+
+```python
+from pyoq.query import insert_into
+
+created = (
+    insert_into(USERS, USER_ID, USER_NAME).values(100, "Hermann").values(101, "Alfred")
+)
+```
+
+The selected columns determine the exact positional types and arity that
+`values()` accepts, so a wrong order, type, or count is a static error in both
+strict type checkers. Values become bound parameters; typed expressions may be
+passed where a computed value is required.
+
+Generated immutable insert values feed the same statement through
+`values_many()`:
+
+```python
+statement = insert_into(USERS, USER_ID, USER_NAME).values_many(
+    (User.builder().id(100).name("Hermann").build(),)
+)
+```
+
+`values_many()` maps each generated value onto the selected columns by the
+field name recorded on the generated column descriptor. A value that leaves a
+selected column unset is rejected, because a selected column always requires a
+value. Handwritten `field()` columns carry no generated field name and are
+therefore positional only.
+
+Generated columns cannot be written. Selecting one, or selecting a column the
+schema marks read-only, fails with a `QueryValidationError` before any SQL is
+produced.
+
+`returning()` changes the terminal result type and produces a statement that
+result operations can read:
+
+```python
+row = database.one(
+    insert_into(USERS, USER_NAME).values("Hermann").returning(USER_ID, USER_NAME)
+)
+```
+
+RETURNING requires SQLite 3.35 or later. It is governed by the `returning`
+compiler capability, and disabling that capability makes the clause fail closed
+with `UnsupportedQueryError`.
+
+Statements without a returning clause execute through `execute()`, which
+reports the affected row count and the last inserted row identifier:
+
+```python
+result = database.execute(insert_into(USERS, USER_ID, USER_NAME).values(100, "Hermann"))
+```
+
+Run the typed INSERT example with:
+
+```console
+python examples/typed_inserts.py
+```
+
+## Typed UPDATE and DELETE
+
+`update(table)` and `delete_from(table)` build the same kind of immutable
+statement. Assignments are typed by their column, and values become bound
+parameters.
+
+```python
+from pyoq.query import delete_from, update
+
+renamed = update(USERS).set(USER_NAME, "Ada Lovelace").where(USER_ID.eq(1))
+removed = delete_from(USERS).where(USER_ID.eq(1))
+```
+
+A statement that would touch every row must say so. An UPDATE or DELETE with no
+WHERE condition fails with a `CompilationError` before any SQL reaches the
+database, so a forgotten condition cannot rewrite or empty a table:
+
+```python
+update(USERS).set(USER_ACTIVE, False).all_rows()
+delete_from(USERS).all_rows()
+```
+
+`all_rows()` and `where()` exclude each other. Defining both, or defining either
+twice, raises `QueryStateError`.
+
+Generated immutable update values assign only the fields they actually set:
+
+```python
+statement = update(USERS).set_values(
+    User.update_builder().name("Ada").build(),
+    USER_NAME,
+    USER_ACTIVE,
+)
+```
+
+Unset fields are skipped, which is what makes an update value a partial update.
+A value that leaves every named column unset is rejected, because that statement
+would assign nothing. As with inserts, generated columns and read-only columns
+cannot be assigned.
+
+`returning()` is available on every write form and produces the same typed
+terminal statement:
+
+```python
+row = database.one(
+    delete_from(USERS).where(USER_ID.eq(1)).returning(USER_ID, USER_NAME)
+)
+```
+
+Run the typed UPDATE and DELETE example with:
+
+```console
+python examples/typed_updates.py
+```
+
+## Conflict resolution
+
+An INSERT can resolve a uniqueness conflict instead of failing. The conflict
+target names the columns whose constraint is being resolved:
+
+```python
+from pyoq.query import excluded, insert_into
+
+ignored = (
+    insert_into(USERS, USER_ID, USER_NAME)
+    .values(1, "Ada")
+    .on_conflict_do_nothing(USER_ID)
+)
+
+merged = (
+    insert_into(USERS, USER_ID, USER_NAME)
+    .values(1, "Ada Lovelace")
+    .on_conflict_do_update(USER_ID)
+    .set(USER_NAME, excluded(USER_NAME))
+)
+```
+
+`excluded()` refers to the row the statement tried to insert, keeping the value
+type of the column it names. It is how a conflicting update reads the new value
+rather than the stored one.
+
+A conflict target must be one of the columns the statement inserts, so a typo or
+a column from another table fails with a `QueryValidationError` rather than
+producing a statement that never matches. `on_conflict_do_nothing()` may omit the
+target, which resolves a conflict on any constraint. Resolution that updates
+always requires a target, because SQLite cannot infer one.
+
+A conflicting update can also filter which rows it touches:
+
+```python
+merged = (
+    insert_into(USERS, USER_ID, USER_NAME)
+    .values(1, "Ada Lovelace")
+    .on_conflict_do_update(USER_ID)
+    .set(USER_NAME, excluded(USER_NAME))
+    .where(USER_NAME.ne("locked"))
+)
+```
+
+Conflict resolution can be defined once per statement, and an update resolution
+that assigns nothing fails at compilation. The `upsert` compiler capability
+disables the whole feature and makes it fail closed with
+`UnsupportedQueryError`.
+
+Run the conflict resolution example with:
+
+```console
+python examples/typed_upserts.py
+```
+
+## Bulk and multi-operation writes
+
+SQLite binds a bounded number of parameters per statement, so a large multi-row
+insert cannot be one statement. The bulk planner splits it into the fewest
+statements that each stay inside the budget, and it does so without executing
+anything:
+
+```python
+plan = database.plan_bulk(statement)
+
+plan.statement_count
+plan.input_rows
+plan.row_counts
+```
+
+Chunks are contiguous and cover every input row exactly once, so each planned
+statement maps back to a known range of input rows through `start` and
+`length`. That is the correlation the plan guarantees. SQLite does not define
+the row order a `RETURNING` clause produces, so returned rows correlate to a
+chunk rather than to an individual input row.
+
+Planning accounts for the parameters a statement spends outside its rows.
+Conflict assignments, conflict conditions, and returning projections reserve
+their share of the budget before rows are packed, and rows whose values are not
+plain bound parameters are measured individually rather than assumed.
+
+```python
+result = database.execute_bulk(statement)
+
+result.rows_affected
+result.statement_count
+result.row_counts
+```
+
+A bulk write is only atomic inside a transaction. Executed directly, each
+planned statement commits on its own, so a failure in a later chunk leaves the
+earlier chunks applied. Wrap the call in a transaction when the whole batch must
+succeed or fail together:
+
+```python
+with database.transaction() as transaction:
+    transaction.execute_bulk(statement)
+```
+
+`many_bulk()` runs the same plan and returns the returned rows from every chunk.
+
+Ordered multi-operation execution runs a sequence of write statements in order
+under an explicit budget, so an unbounded batch cannot be submitted by accident:
+
+```python
+from pyoq.query.execution import OperationBudget
+
+results = database.execute_all(
+    (first_statement, second_statement),
+    budget=OperationBudget(maximum_operations=8),
+)
+```
+
+An empty sequence executes nothing and returns an empty result tuple. Execution
+stops at the first failure, and the same atomicity rule applies: use a
+transaction when the earlier operations must not survive a later failure.
+
+Run the bulk write example with:
+
+```console
+python examples/bulk_writes.py
+```
+
+## SQLite compilation
+
+`SQLiteCompiler` renders a query into immutable SQL text and a separate ordered
+parameter tuple. Python values never enter the SQL string. Pagination values
+use the same parameter path, and sensitive bound values are recorded by their
+zero-based parameter positions.
+
+```python
+from pyoq.query import bind, select
+from pyoq.query.sqlite import SQLiteCompiler
+
+compiled = SQLiteCompiler().compile(
+    select(USER_ID, USER_NAME)
+    .from_(USERS)
+    .where(USER_ID.gt(bind(100, sensitive=True)))
+    .limit(20)
+)
+
+assert compiled.parameters == (100, 20)
+assert compiled.sensitive_parameter_indexes == frozenset({0})
+```
+
+Identifiers are quoted component by component. Operator expressions are
+parenthesized deterministically, null equality is rendered with `IS NULL`, and
+empty membership predicates become constant boolean expressions. Derived
+tables, common tables, joins, aggregates, and compound queries preserve the
+textual order of their parameters.
+
+```mermaid
+flowchart LR
+    Query[Immutable query tree] --> Capability[SQLite capability checks]
+    Capability --> Compiler[SQLite compiler]
+    Compiler --> SQL[Quoted SQL structure]
+    Compiler --> Parameters[Ordered parameter tuple]
+    Compiler --> Sensitive[Sensitive parameter positions]
+```
+
+Compiler capabilities are explicit and immutable. RIGHT JOIN and FULL JOIN are
+disabled by default because support depends on the SQLite runtime. Recursive
+common tables, explicit null ordering, raw expressions, write returning,
+conflict resolution, and the maximum parameter budget can also be restricted. A query requiring a
+disabled or unrepresentable feature fails with an exception from `pyoq.errors`.
+
+The compiler also renders write statements. An INSERT compiles into one
+statement whose rows share the same quoted column list, an UPDATE renders its
+assignments in the order they were added, and every value in either form is a
+bound parameter. An UPDATE or DELETE that has neither a WHERE condition nor an
+explicit full-table opt-in fails here rather than reaching the database.
+
+Temporal arithmetic is not translated into SQLite numeric operators. Use an
+explicit typed raw expression when the intended SQLite date function is known.
+Raw expression text cannot introduce bind markers, statement separators, or
+SQL comment syntax. Catalog-qualified identifiers also fail because SQLite has
+no matching catalog level.
+
+Run the compiler example with:
+
+```console
+python examples/sqlite_compilation.py
+```
+
+## Synchronous SQLite execution
+
+`SQLitePool` provides bounded, exclusive connection leases. One checked-out
+connection belongs to one caller until the lease exits. Application exceptions
+return healthy connections, while failures that indicate a broken connection
+discard only that connection. Checkout waits are bounded by policy.
+
+```python
+from pyoq.query import select
+from pyoq.query.sqlite import (
+    SQLiteConnectionFactory,
+    SQLiteExecutor,
+    SQLitePool,
+    SQLitePoolPolicy,
+)
+
+pool = SQLitePool(
+    SQLiteConnectionFactory("application.sqlite"),
+    SQLitePoolPolicy(maximum_size=10, checkout_timeout=5.0),
+)
+db = SQLiteExecutor(pool)
+
+rows: list[tuple[int, str]] = db.many(
+    select(USER_ID, USER_NAME).from_(USERS).order_by(USER_ID)
+)
+selected: tuple[int, str] = db.one(
+    select(USER_ID, USER_NAME).from_(USERS).where(USER_ID.eq(1))
+)
+name: str = db.scalar(select(USER_NAME).from_(USERS).where(USER_ID.eq(1)))
+
+pool.close()
+```
+
+`one` requires exactly one row, `one_or_none` accepts zero or one, `many`
+returns all rows, and `scalar` requires a single projected value. Their return
+types follow the query projection. `execute` accepts an immutable compiled
+statement and returns affected-row metadata. Insert identifiers are populated
+only for statements carrying explicit insert classification, which prevents
+stale driver metadata from escaping.
+
+SQLite-compatible scalar parameters are adapted before a connection is
+checked out. Text, numbers, bytes, temporal values, decimal values, UUIDs,
+enums, memory views, and JSON containers have deterministic representations.
+Unsupported values fail with `ParameterBindingError` before pool capacity is
+used.
+
+```mermaid
+flowchart LR
+    Query[Typed query] --> Compiler[SQLite compiler]
+    Compiler --> Adapter[Parameter adapter]
+    Adapter --> Pool[Exclusive pool lease]
+    Pool --> Driver[SQLite driver]
+    Driver --> Cardinality[Typed result policy]
+    Cardinality --> Result[Row, scalar, or execute result]
+```
+
+Run the execution example with:
+
+```console
+python examples/sqlite_execution.py
+```
+
+## Transactions and bounded streaming
+
+`SQLiteTransaction` owns one pool lease for its complete context. Normal exit
+commits, while every exception derived from `BaseException` rolls back. A
+transaction can only be used by its owner thread, and only its innermost active
+scope may issue work. Nested scopes use generated savepoint identifiers.
+
+```python
+from pyoq.query.sqlite import TransactionMode
+
+with db.transaction(TransactionMode.IMMEDIATE) as transaction:
+    transaction.execute(FIRST_INSERT)
+
+    try:
+        with transaction.savepoint() as nested:
+            nested.execute(OPTIONAL_INSERT)
+            raise ValueError("discard optional work")
+    except ValueError:
+        pass
+
+    transaction.execute(FINAL_INSERT)
+```
+
+Statement failures that leave the SQLite connection healthy can be handled
+inside the transaction. Connection failures make the complete transaction
+unusable. A commit or savepoint failure discards the connection because its
+outcome cannot be assumed safely.
+
+`stream` returns a typed iterator that fetches at most the configured batch
+size. It checks out no connection until it is entered or iterated, returns the
+lease automatically on exhaustion, and closes on iteration failure. Use its
+context manager whenever iteration may stop early.
+
+```python
+from pyoq.query.execution import CancellationToken, ExecutionControl
+from pyoq.query.sqlite import SQLiteStreamPolicy
+
+token = CancellationToken()
+policy = SQLiteStreamPolicy(
+    batch_size=128,
+    control=ExecutionControl(
+        timeout=2.5,
+        cancellation_token=token,
+        progress_steps=500,
+    ),
+)
+
+with db.stream(
+    select(USER_ID, USER_NAME).from_(USERS).order_by(USER_ID),
+    policy=policy,
+) as rows:
+    for user_id, user_name in rows:
+        consume(user_id, user_name)
+```
+
+Timeout and cancellation checks run before execution, through SQLite progress
+callbacks, and between streamed rows. `QueryTimeoutError` and
+`QueryCancelledError` are distinct typed failures. Cancellation is cooperative:
+another thread may call `token.cancel()`, but it must not use the owned
+connection or consume the row stream.
+
+```mermaid
+flowchart TD
+    Executor[SQLiteExecutor] --> Lease[Exclusive pool lease]
+    Lease --> Transaction[Owned transaction]
+    Transaction --> Savepoint[Nested savepoint]
+    Transaction --> Stream[Bounded row stream]
+    Stream --> Batch[At most batch_size buffered rows]
+    Control[Timeout or cancellation] --> Progress[SQLite progress callback]
+    Progress --> Cleanup[Cursor, handler, transaction, and lease cleanup]
+    Batch --> Cleanup
+    Savepoint --> Cleanup
+```
+
+Run the transaction and streaming example with:
+
+```console
+python examples/sqlite_transactions.py
+```
+
+## Command line
+
+The package installs the `pyoq` command and also supports `python -m pyoq`.
+
+```console
+pyoq --version
+pyoq generate --project-root .
+pyoq generate --project-root . --check
+pyoq generate --project-root . --dry-run
+pyoq inspect --project-root .
+```
+
+Both database commands accept `--config` for a configuration file below the
+project root and `--profile` for explicit profile selection. `--check` and
+`--dry-run` are mutually exclusive.
+
+Command exit codes are stable:
+
+- `0`: success
+- `1`: command service failure
+- `2`: invalid command usage
+- `3`: configuration failure
+- `4`: requested operation is unavailable
+
+SQLite generation and inspection are available through the default command
+services. Inspection reads tables, views, columns, defaults, generated values,
+primary and unique keys, indexes, CHECK constraints, and foreign-key
+relationships. Metadata is normalized before rendering, and the connection is
+closed before generated files are checked or written.
+
+Generation writes six deterministic, fully typed modules and an ownership
+manifest into `codegen-directory`. Repeating a write against an unchanged
+database leaves the generated directory untouched. `--check` fails on drift,
+while `--dry-run` reports planned creation, update, removal, and conflict counts
+without changing files. Generation runs against SQLite, PostgreSQL, and
+MySQL.
+
+Run the complete SQLite reflection and generation example with:
+
+```console
+python examples/sqlite_generation.py
+```
+
+## Performance contract
+
+Run the native and portable kernel comparison from an editable native build:
+
+```console
+python -m benchmarks.performance_contract
+```
+
+The command checks relative throughput, traced peak memory, retained
+allocations, and native artifact size. It exits unsuccessfully when a measured
+kernel exceeds its calibrated budget.
+
+## Distribution checks
+
+Build the native wheel with the release backend:
+
+```console
+maturin build --release --locked --out wheelhouse
+```
+
+Build the universal portable wheel without compiling the native extension:
+
+```console
+hatchling build -t wheel -d wheelhouse
+```
+
+When both artifact kinds are present, validate their tags, native contents,
+license, type information, and shared package metadata:
+
+```console
+python -m scripts.verify_wheels wheelhouse
+```
+
+Continuous integration builds native wheels for glibc and musl Linux on x86-64
+and ARM64, macOS on Apple Silicon and x86-64, and Windows on x86-64 and ARM64.
+The stable native ABI is installed and imported on every supported Python
+version. The universal wheel is also installed through an incompatible-platform
+resolver test to prove that it does not require Rust.
+
+## Design goals
+
+- Fully typed public APIs compatible with mypy and Pyright strict modes.
+- Database-first schema generation with deterministic output.
+- Explicit, immutable SQL query models and bound parameters.
+- PostgreSQL, MySQL, and SQLite support.
+- Django, FastAPI, and Sanic integrations.
+- Connection pooling, transaction safety, and observable query execution.
+- Fetch planning that prevents hidden I/O and N+1 query behavior.
+- Native hot paths with measured throughput, allocation, and peak-memory gates.
+
+## Supported Python
+
+PyOQ requires Python 3.11 or newer.
+
+## License
+
+PyOQ is licensed under the Mozilla Public License 2.0. Applications may use it
+commercially, including as part of larger proprietary products. Changes to
+covered PyOQ source files remain subject to the MPL terms. See `LICENSE` for the
+complete terms.