diff --git a/CHANGELOG.md b/CHANGELOG.md
deleted file mode 100644
index 1d3ffe5..0000000
--- a/CHANGELOG.md
+++ /dev/null
@@ -1,9 +0,0 @@
-# Changelog
-
-## Unreleased
-- Require PHP 8.3+ and modernize PHPUnit to 10.x.
-- Add PostgreSQL compatibility alongside MySQL/MariaDB.
-- Improve identifier quoting and driver-specific schema discovery.
-- Update tests to support configurable DB backends.
-- Add GitHub Actions CI matrix for PHP and database drivers.
-- Add automatic CalVer tagging and GitHub releases for merged PRs.
diff --git a/ROADMAP.md b/ROADMAP.md
index 040acd1..a0e6057 100644
--- a/ROADMAP.md
+++ b/ROADMAP.md
@@ -1,12 +1,20 @@
 # Roadmap
 
-This roadmap turns the current improvement analysis into a concrete execution plan for `freshsauce/model`.
+This roadmap tracks the improvement work for `freshsauce/model`.
 
-The sequencing is intentional:
+Current status:
+
+1. Phase 1 is complete.
+2. Phase 2 is complete.
+3. Phase 3 is complete.
+4. Phase 4 remains optional and has not been started.
+
+The sequencing remains intentional:
 
 1. Fix correctness issues before expanding the API.
 2. Improve developer ergonomics without turning the library into a heavyweight ORM.
-3. Add optional features only where they preserve the package's lightweight position.
+3. Tighten quality and portability before considering broader feature growth.
+4. Add optional features only where they preserve the package's lightweight position.
 
 ## Principles
 
@@ -17,249 +25,33 @@ The sequencing is intentional:
 
 ## Phase 1: Core correctness and safety
 
-Goal: remove known edge-case bugs and make failure modes explicit.
-
-Priority: high
-
-### Milestone 1.1: Serialization works correctly
-
-Problem:
-`Model::__sleep()` returns table field names, but model state lives in `$data` and `$dirty`. Serializing a model currently emits warnings and drops state.
-
-Tasks:
-
-- Replace `__sleep()` with `__serialize()` and `__unserialize()`.
-- Preserve hydrated values and dirty-state behavior after unserialization.
-- Decide whether deserialized models should retain dirty flags or reset to clean.
-- Add PHPUnit coverage for round-trip serialization of new and persisted models.
-
-Acceptance criteria:
-
-- `serialize($model)` produces no warnings.
-- `unserialize(serialize($model))` preserves field values.
-- Dirty-state behavior after round-trip is documented and tested.
-
-### Milestone 1.2: Persisted-state detection is explicit
-
-Problem:
-`save()` uses truthiness on the primary key. A value like `0` is treated as "new" and triggers insert instead of update.
-
-Tasks:
-
-- Introduce a dedicated persisted-state check based on `null` rather than truthiness.
-- Review insert/update/delete behavior for zero-like primary key values.
-- Add tests for integer `0`, string `'0'`, and non-default primary key names.
-
-Acceptance criteria:
-
-- `save()` updates when a record has a zero-like primary key value.
-- Insert behavior remains unchanged for `null` primary keys.
-
-### Milestone 1.3: Dynamic finder failures are model-level errors
-
-Problem:
-Unknown dynamic fields fall through to raw SQL execution and surface as PDO errors instead of clear model exceptions.
-
-Tasks:
-
-- Make `resolveFieldName()` fail fast when a field does not map to a real column.
-- Add a dedicated exception for unknown fields or invalid dynamic methods.
-- Add tests for invalid `findBy...`, `findOneBy...`, and `countBy...` calls.
+Status: completed
 
-Acceptance criteria:
+Summary:
 
-- Invalid dynamic finders throw a predictable library exception before query execution.
-- Error messages identify the requested field and model class.
-
-### Milestone 1.4: Empty-array query behavior is defined
-
-Problem:
-Helpers that build `IN (...)` clauses do not define behavior for empty arrays.
-
-Tasks:
-
-- Define expected behavior for empty-array matches across:
-  - `findBy...`
-  - `findOneBy...`
-  - `firstBy...`
-  - `lastBy...`
-  - `countBy...`
-  - `fetchAllWhereMatchingSingleField()`
-  - `fetchOneWhereMatchingSingleField()`
-- Implement that behavior without generating invalid SQL.
-- Add regression tests for each public entry point.
-
-Acceptance criteria:
-
-- Empty arrays never produce invalid SQL.
-- Collection methods return empty results.
-- Singular methods return `null`.
-- Count methods return `0`.
-
-### Milestone 1.5: Replace generic exceptions with library exceptions
-
-Problem:
-The model currently throws generic `\Exception` in many places, which makes calling code and tests less precise.
-
-Tasks:
-
-- Introduce a small exception hierarchy under `Freshsauce\Model\Exception\`.
-- Replace generic throws for:
-  - missing connection
-  - unknown field
-  - invalid dynamic method
-  - missing data access
-  - identifier quoting setup failures
-- Keep exception names narrow and practical.
-
-Acceptance criteria:
-
-- Core failure modes throw specific exception classes.
-- Existing messages remain readable.
-- Public docs mention the main exception types users should expect.
+- Fixed serialization, zero-like primary key handling, invalid dynamic finder failures, and empty-array query behavior.
+- Replaced generic exceptions with a small library exception hierarchy.
+- Added regression coverage for the above edge cases.
 
 ## Phase 2: API ergonomics and typing
 
-Goal: make the library easier to use correctly while keeping the current lightweight style.
-
-Priority: medium
-
-### Milestone 2.1: Validation becomes instance-aware
-
-Problem:
-`validate()` is declared static, but it is invoked from instance writes. That makes record-aware validation awkward.
-
-Tasks:
-
-- Change validation to an instance hook, or introduce `validateForInsert()` and `validateForUpdate()` instance hooks.
-- Preserve backward compatibility where practical, or provide a clear migration path.
-- Add tests covering validation against current field values.
-
-Acceptance criteria:
-
-- Validation can inspect instance state directly.
-- Validation behavior for insert vs update is documented.
+Status: completed
 
-### Milestone 2.2: Strict field handling is available
+Summary:
 
-Problem:
-Unknown fields can be assigned silently via `__set()`, but are ignored during persistence. That hides typos.
-
-Tasks:
-
-- Add an optional strict mode that rejects assignment to unknown fields.
-- Decide whether strict mode is global, per model, or opt-in at runtime.
-- Keep the current permissive mode available for legacy integrations.
-- Add tests for strict and permissive behavior.
-
-Acceptance criteria:
-
-- Strict mode raises a clear exception on unknown field assignment.
-- Default behavior remains stable unless the user opts in.
-
-### Milestone 2.3: Add focused query helpers
-
-Problem:
-Many common query patterns require manual SQL fragments, which works but is unnecessarily error-prone.
-
-Tasks:
-
-- Add a small set of helpers with clear value:
-  - `exists()`
-  - `existsWhere()`
-  - `pluck(string $field, ... )`
-  - `orderBy(string $field, string $direction)` via new fetch helpers, not a full query builder
-  - `limit(int $n)` support in helper methods where it keeps the API simple
-- Avoid introducing a chainable query-builder unless a later need is proven.
-
-Acceptance criteria:
-
-- Common "check existence", "single column list", and "ordered fetch" cases need less handwritten SQL.
-- The API remains smaller than a full query-builder abstraction.
-
-### Milestone 2.4: Tighten types and static analysis
-
-Problem:
-The library runs on PHP 8.3+ but still carries loose typing in several public and protected methods.
-
-Tasks:
-
-- Add `declare(strict_types=1);` to source and tests.
-- Add explicit parameter and return types where missing.
-- Improve PHPDoc for dynamic methods and arrays.
-- Reduce `phpstan.neon` ignores where feasible.
-
-Acceptance criteria:
-
-- PHPStan remains green at the current level.
-- Public APIs are more explicit and easier to consume from IDEs.
-
-### Milestone 2.5: Refresh documentation around modern usage
-
-Problem:
-The docs still present deprecated snake_case dynamic methods in examples and do not explain newer behavior clearly enough.
-
-Tasks:
-
-- Update `README.md` and `EXAMPLE.md` to lead with camelCase dynamic methods only.
-- Add migration notes for deprecated snake_case methods.
-- Document strict mode, validation hooks, and exception behavior once shipped.
-- Add a short "when to use this library" section to reinforce scope boundaries.
-
-Acceptance criteria:
-
-- Public docs reflect the preferred API.
-- Deprecated behavior is documented as transitional, not primary.
+- Added instance-aware validation hooks with legacy compatibility.
+- Added optional strict field handling and focused query helpers.
+- Tightened typing, static analysis, and public documentation around the preferred API.
 
 ## Phase 3: Quality, portability, and maintenance
 
-Goal: make the library easier to maintain and safer across supported databases.
-
-Priority: medium
-
-### Milestone 3.1: Expand edge-case test coverage
-
-Tasks:
-
-- Add regression tests for every Phase 1 fix.
-- Add tests for multiple model subclasses sharing and isolating connections.
-- Add tests for schema-qualified PostgreSQL table names.
-- Add tests for custom primary key column names.
-- Add tests for timestamp opt-out behavior.
-
-Acceptance criteria:
-
-- New fixes are guarded by tests.
-- Cross-driver behavior is better documented in test form.
-
-### Milestone 3.2: Review statement and metadata caching behavior
-
-Problem:
-Statement caching is now keyed by connection, which is good, but metadata and caching behavior should stay predictable as the library grows.
-
-Tasks:
-
-- Audit cache invalidation rules for reconnection and subclass-specific connections.
-- Decide whether table column metadata should be refreshable without reconnecting.
-- Add tests around reconnect and metadata refresh behavior.
-
-Acceptance criteria:
-
-- Reconnection cannot leak stale prepared statements.
-- Metadata caching behavior is documented and tested.
-
-### Milestone 3.3: Normalize exception and timestamp behavior across drivers
-
-Tasks:
-
-- Verify `rowCount()` assumptions for update/delete across supported drivers.
-- Review timestamp formatting consistency for MySQL, PostgreSQL, and SQLite.
-- Ensure identifier quoting and schema discovery stay correct for each supported driver.
+Status: completed
 
-Acceptance criteria:
+Summary:
 
-- Driver-specific behavior is explicit where unavoidable.
-- Public docs do not imply unsupported guarantees.
+- Expanded cross-driver integration coverage for connection sharing, custom keys, metadata refresh, timestamp behavior, and PostgreSQL schema-qualified tables.
+- Added `refreshTableMetadata()` and made UTC timestamp behavior explicit.
+- Normalized no-op update handling while preserving single-row primary key update expectations.
 
 ## Phase 4: Optional feature expansion
 
@@ -267,7 +59,7 @@ Goal: add features that help real applications, but only if they fit the package
 
 Priority: lower
 
-These should only start after Phases 1 and 2 are complete.
+Phases 1 through 3 are complete, so this is now the remaining backlog.
 
 ### Candidate 4.1: Transaction helpers
 
@@ -316,22 +108,17 @@ Recommendation:
 
 If this work is split into GitHub issues, the most practical order is:
 
-1. Replace `__sleep()` with proper serialization support.
-2. Fix zero-like primary key handling in `save()`.
-3. Make unknown dynamic fields fail fast.
-4. Define empty-array query behavior.
-5. Introduce library exception classes.
-6. Expand regression tests for the above.
-7. Convert validation to instance-aware hooks.
-8. Add strict field mode.
-9. Add focused query helpers.
-10. Tighten typing and refresh docs.
+1. Add transaction helpers.
+2. Add configurable timestamp column support.
+3. Add attribute casting.
+4. Re-evaluate whether composite keys or relationship support are warranted.
 
 ## Suggested release strategy
 
-- Release 1: all Phase 1 fixes plus regression tests.
-- Release 2: validation, strict mode, typed API cleanup, and documentation refresh.
-- Release 3: selected optional helpers such as transactions, timestamp configuration, and casting.
+- Release 1: Phase 1 correctness and exception work. Shipped.
+- Release 2: Phase 2 ergonomics, typing, and documentation updates. Shipped.
+- Release 3: Phase 3 portability and maintenance hardening. Shipped.
+- Release 4: optional feature work only if it still fits the package scope.
 
 ## Out of scope unless demand changes