Enforce Oxford spelling

.github/workflows/ci.yml

@@ -50,6 +50,8 @@ jobs:
         with:
           python-version: '3.13'
           enable-cache: true
+      - name: Spelling
+        run: make spelling
       - name: Workflow contract tests
         run: make test-workflow-contracts
       - name: Test

.gitignore

@@ -2,3 +2,9 @@ target/
 **/*.rs.bk
 .memdb/
 __pycache__/
+.coverage
+.pytest_cache/
+.uv-cache/
+.uv-tools/
+.typos-oxendict-base.json
+.typos-oxendict-base.toml

.rules/python-exception-design-raising-handling-and-logging.md

@@ -90,7 +90,7 @@ def reciprocal(n: float) -> float:
         return result
 ```
 
-`else` emphasises the happy path and avoids odd control‑flow within `try`
+`else` emphasizes the happy path and avoids odd control‑flow within `try`
 blocks.
 
 ## 4) Message construction for raises (EM101/EM102) and logging practice (LOG004/LOG007/LOG009/LOG014/LOG015, TRY401)
@@ -110,7 +110,7 @@ raise RuntimeError(msg)
 EM101/EM102 prefer a single message object; this reduces duplication and
 clarifies intent.
 
-### Logging: parameterised messages, module loggers, correct APIs
+### Logging: parameterized messages, module loggers, correct APIs
 
 ```python
 import logging

.rules/python-pyproject.md

@@ -228,7 +228,7 @@ package = true
      Reddit[^5])
    - `description`, `readme`, `requires-python`: provide clarity about the
      project and help tools like PyPI. (Python Packaging[^4], Reddit[^5])
-   - `license`, `authors`, `keywords`, `classifiers`: standardised metadata,
+   - `license`, `authors`, `keywords`, `classifiers`: standardized metadata,
      which improves discoverability. (Python Packaging[^4], Reddit[^5])
    - `dependencies`: runtime requirements, expressed in PEP 508 syntax.
      (Astral Docs[^1], RidgeRun.ai[^2])

.rules/python-typing.md

@@ -8,7 +8,7 @@
 
 Use `Enum` for fixed sets of related constants. Use `enum.auto()` to avoid
 repeating values manually. Use `IntEnum` or `StrEnum` when interoperability
-with integers or strings is required (e.g. for database or JSON serialisation).
+with integers or strings is required (e.g. for database or JSON serialization).
 
 ```python
 import enum

AGENTS.md

@@ -28,7 +28,7 @@
   examples demonstrating the usage and outcome of the function. Test
   documentation should omit examples where the example serves only to reiterate
   the test logic.
-- **Keep file size managable.** No single code file may be longer than 400
+- **Keep file size manageable.** No single code file may be longer than 400
   lines. Long switch statements or dispatch tables should be broken up by
   feature and constituents colocated with targets. Large blocks of test data
   should be moved to external data files.
@@ -59,7 +59,7 @@
   - Where a bug is being fixed, a unittest has been provided demonstrating the
     behaviour being corrected both to validate the fix and to guard against
     regression.
-  - Passes all relevant unit and behavioral tests according to the guidelines
+  - Passes all relevant unit and behavioural tests according to the guidelines
     above. (Use `make test` to verify).
   - Passes lint checks. (Use `make lint` to verify).
   - Adheres to formatting standards tested using a formatting validator. (Use
@@ -106,7 +106,7 @@
 - **Separate Atomic Refactors:** If refactoring is deemed necessary:
   - Perform the refactoring as a **separate, atomic commit** *after* the
     functional change commit.
-  - Ensure the refactoring adheres to the testing guidelines (behavioral tests
+  - Ensure the refactoring adheres to the testing guidelines (behavioural tests
     pass before and after, unit tests added for new units).
   - Ensure the refactoring commit itself passes all quality gates.
 
@@ -173,7 +173,7 @@ project:
 - Lint rule suppressions must be tightly scoped and include a clear reason.
 - Prefer `expect` over `allow`.
 - Use `rstest` fixtures for shared setup.
-- Replace duplicated tests with `#[rstest(...)]` parameterised cases.
+- Replace duplicated tests with `#[rstest(...)]` parameterized cases.
 - Prefer `mockall` for mocks/stubs.
 - Prefer `.expect()` over `.unwrap()`.
 - Use `concat!()` to combine long string literals rather than escaping newlines
@@ -199,7 +199,7 @@ project:
   `into_inner()` and `to_path_buf()`; avoid attempting
   `impl From<Wrapper> for PathBuf` because of the orphan rule. Prefer explicit
   tuple structs whenever bespoke validation or tailored trait surfaces are
-  required, customising `Deref`, `AsRef`, and `TryFrom` per type. Use
+  required, customizing `Deref`, `AsRef`, and `TryFrom` per type. Use
   `the-newtype` when defining traits and needing blanket implementations that
   apply across wrappers satisfying `Newtype + AsRef/AsMut<Inner>`, or when
   establishing a coherent internal convention that keeps trait forwarding
@@ -212,7 +212,7 @@ project:
 - Write unit and behavioural tests for new functionality. Run both before and
   after making any change.
 - Use `rstest` fixtures for shared setup.
-- Replace duplicated tests with `#[rstest(...)]` parameterised cases.
+- Replace duplicated tests with `#[rstest(...)]` parameterized cases.
 - Prefer `mockall` for mocks/stubs.
 - Mock non-deterministic dependencies (e.g., environment variables and the
   system clock) using dependency injection with the `mockable` crate (traits

Makefile

@@ -1,17 +1,22 @@
-.PHONY: help all clean test test-workflow-contracts build release lint typecheck fmt check-fmt markdownlint nixie bless
+.PHONY: help all clean test test-workflow-contracts build release lint typecheck \
+	fmt check-fmt markdownlint nixie spelling spelling-helper-test bless
 
-APP ?= lag-complexity
+APP ?= lagc
 CARGO ?= cargo
 BUILD_JOBS ?=
 CLIPPY_FLAGS ?= --all-targets --all-features -- -D warnings
 MDLINT ?= markdownlint-cli2
 NIXIE ?= nixie
 WHITAKER ?= whitaker
+UV ?= uv
+UV_ENV = UV_CACHE_DIR=.uv-cache UV_TOOL_DIR=.uv-tools
+RUFF_VERSION ?= 0.15.12
+TYPOS_VERSION ?= 1.48.0
 
 build: target/debug/$(APP) ## Build debug binary
 release: target/release/$(APP) ## Build release binary
 
-all: release ## Default target builds release binary
+all: release spelling ## Build the release binary and enforce spelling
 
 clean: ## Remove build artifacts
 	$(CARGO) clean
@@ -39,9 +44,31 @@ fmt: ## Format Rust and Markdown sources
 check-fmt: ## Verify formatting
 	$(CARGO) fmt --all -- --check
 
-markdownlint: ## Lint Markdown files
+markdownlint: spelling ## Lint Markdown files and enforce spelling
 	find . -type f -name '*.md' -not -path './target/*' -print0 | xargs -0 $(MDLINT)
 
+spelling: spelling-helper-test ## Enforce en-GB-oxendict spelling in Markdown prose
+	@$(UV_ENV) $(UV) run scripts/generate_typos_config.py
+	@git ls-files -z '*.md' | \
+		xargs -0 -r env $(UV_ENV) $(UV) tool run typos@$(TYPOS_VERSION) \
+		--config typos.toml --force-exclude
+
+spelling-helper-test: ## Validate the shared spelling-policy integration
+	@$(UV_ENV) $(UV) tool run ruff@$(RUFF_VERSION) format --isolated \
+		--target-version py313 --check scripts/generate_typos_config.py \
+		scripts/typos_rollout.py scripts/typos_rollout_cache.py \
+		scripts/tests/test_typos_rollout.py
+	@$(UV_ENV) $(UV) tool run ruff@$(RUFF_VERSION) check --isolated \
+		--target-version py313 scripts/generate_typos_config.py \
+		scripts/typos_rollout.py scripts/typos_rollout_cache.py \
+		scripts/tests/test_typos_rollout.py
+	@PYTHONPATH=scripts $(UV_ENV) $(UV) run --no-project --python 3.13 \
+		--with pytest==9.0.2 --with pytest-cov==7.0.0 \
+		python -m pytest scripts/tests/test_typos_rollout.py \
+		-c /dev/null --rootdir=. -p no:cacheprovider \
+		--cov=generate_typos_config --cov=typos_rollout \
+		--cov=typos_rollout_cache --cov-fail-under=90
+
 nixie: ## Validate Mermaid diagrams
 	# CI currently requires --no-sandbox; remove once nixie supports
 	# environment variable control for this option

docs/complexity-antipatterns-and-refactoring-strategies.md

@@ -2,7 +2,7 @@
 
 ## I. Introduction: The Challenge of Software Complexity
 
-Software development is an inherently complex endeavor. As systems evolve and
+Software development is an inherently complex endeavour. As systems evolve and
 features are added, the intricacy of the codebase tends to increase, often
 leading to challenges in maintenance, scalability, and developer productivity.
 "Any fool can write code that a computer can understand. Good programmers write
@@ -95,7 +95,7 @@ Cognitive Complexity is incremented based on three main rules 8:
 3. **Shorthand Discount:** Structures that allow multiple statements to be read
    as a single unit (e.g., a well-named method call) do not incur the same
    penalties as the raw statements they encapsulate. Method calls are generally
-   "free" in terms of cognitive complexity, as a well-chosen name summarises
+   "free" in terms of cognitive complexity, as a well-chosen name summarizes
    the underlying logic, allowing readers to grasp the high-level view before
    diving into details. However, recursive calls do increment the score.8
 
@@ -110,7 +110,7 @@ Thresholds and Implications:
 
 Code with high Cognitive Complexity is harder to read, understand, test, and
 modify.8 SonarQube, for example, raises issues when a function's Cognitive
-Complexity exceeds a certain threshold, signaling that the code should likely
+Complexity exceeds a certain threshold, signalling that the code should likely
 be refactored into smaller, more manageable pieces.8 The primary impact of high
 Cognitive Complexity is a slowdown in development and an increase in
 maintenance costs.8
@@ -338,7 +338,7 @@ perform auto-refactoring for certain languages.6
 
 ### C. Red Flags Portending the Bumpy Road
 
-Recognising early warning signs can prevent minor complexity issues from
+Recognizing early warning signs can prevent minor complexity issues from
 escalating into full-blown Bumpy Roads.
 
 1. **Increasing Cognitive Complexity Scores:** A rising Cognitive Complexity

docs/developers-guide.md

@@ -4,6 +4,14 @@ This guide records development practices that are specific to maintaining this
 repository. Follow the project-wide guidance in `AGENTS.md` first, then use
 this guide for workflow automation details.
 
+## Spelling policy
+
+Run `make spelling` to enforce en-GB-oxendict prose spelling. The generated
+`typos.toml` starts from the shared estate dictionary, refreshes its untracked
+local cache only when the authority is newer, and then applies the narrow
+repository policy in `typos.local.toml`. Edit the local policy and regenerate
+the configuration rather than changing generated entries by hand.
+
 ## Continuous Integration workflow
 
 The Continuous Integration (CI) workflow lives in `.github/workflows/ci.yml`.

docs/lag-complexity-function-design.md

@@ -381,7 +381,7 @@ require a wider retrieval strategy.
   and results in a dedicated `VarianceError::Empty`. This guards against silent
   NaNs and ensures caller intent is explicit.
 - **Non-finite Inputs:** `NaN` or infinite values yield `NaN`. Callers should
-  sanitise embeddings if such values are not expected.
+  sanitize embeddings if such values are not expected.
 
 #### Embedding providers
 

docs/model-training-pipeline-design.md

@@ -30,7 +30,7 @@ Key improvements in this design include:
   (training, evaluation, etc.) is executed on dedicated ephemeral cloud
   instances. Prefect tasks use Terraform (OpenTofu distribution) to provision
   the required **spot or on-demand VM instances** (with GPU for training, CPU
-  for export/evaluation) and tear them down afterward. This ensures
+  for export/evaluation) and tear them down afterwards. This ensures
   **cloud-agnostic deployment** – by swapping Terraform modules, the pipeline
   can target AWS, GCP, or other providers uniformly. It also maximizes cost
   efficiency by using spot instances and releasing resources immediately when
@@ -102,7 +102,7 @@ adhere to several core principles in designing the workflow:
   versioned. We use **containerization** (Docker images) and
   infrastructure-as-code to ensure that a given pipeline run can be exactly
   replicated in the future. This is critical for auditing and for guaranteeing
-  consistent model behavior between training and production. The pipeline will
+  consistent model behaviour between training and production. The pipeline will
   incorporate configuration files (or Prefect configuration blocks) for all
   adjustable parameters (dataset version, model architecture, learning rates,
   instance types, etc.), which are stored in version control. By fixing these
@@ -220,7 +220,7 @@ automate their lifecycle:
   2.3). The Terraform provisioning task can request a spot VM (e.g., AWS EC2
   Spot or GCP Preemptible VM) for training; if the request is not fulfilled or
   the instance is reclaimed mid-training, the Prefect flow will handle
-  launching a replacement and resuming training. The **cost modeling** takes
+  launching a replacement and resuming training. The **cost modelling** takes
   into account the expected interruptions and restart overhead, ensuring that
   even with occasional restarts, using spot instances yields net savings. For
   cases where spot is not available or for final runs, the pipeline can easily

docs/model-training-pipeline-roadmap.md

@@ -47,10 +47,10 @@ progress tracking.
   - [ ] 1.1.3.4. Wire alerting (Slack, PagerDuty, email) to Prefect block
         failures and add runbook links for operator response.
 
-### 1.2. Standardise compute procurement and cost controls
+### 1.2. Standardize compute procurement and cost controls
 
 - [ ] Automate OpenTofu modules for g4dn.xlarge, g6.xlarge, and c6i.xlarge
-      instances with provider parameterisation and teardown under 10 minutes.
+      instances with provider parameterization and teardown under 10 minutes.
 - [ ] Validate spot-to-on-demand fallbacks by injecting capacity failures and
       confirming retries succeed within three orchestration attempts.
 - [ ] Publish hourly cost benchmarks for each instance class with variance
@@ -75,7 +75,7 @@ progress tracking.
 
 - [ ] Build ingestion scripts that hydrate `/datasets/raw/` from approved
       sources and emit checksums recorded in metadata manifests.
-- [ ] Create preprocessing jobs that materialise `/datasets/processed/`
+- [ ] Create preprocessing jobs that materialize `/datasets/processed/`
       artefacts with schema validation and rejection workflows for anomalies.
 - [ ] Introduce calibration dataset generation producing 100–500 curated
       samples stored under `/datasets/processed/{dataset}/calibration/`.
@@ -84,7 +84,7 @@ progress tracking.
 
 - [ ] Develop the `OrdinalRegressionTrainer` with cutpoint ordering guarantees
       and unit tests covering boundary cases (monotonicity, identical labels).
-- [ ] Parameterise model head replacement logic to support both BERT-like and
+- [ ] Parameterize model head replacement logic to support both BERT-like and
       ViT architectures, validated on representative text and image datasets.
 - [ ] Achieve ≥90% statement coverage on the training module using `pytest`
       plus fixtures from `rstest`.
@@ -100,7 +100,7 @@ progress tracking.
 - [ ] Capture structured metrics (loss, cutpoints, throughput) and emit them to
       Prefect Orion and the observability stack with comparison dashboards.
 
-## 3. Operationalise export and optimisation
+## 3. Operationalize export and optimization
 
 ### 3.1. Build ONNX export automation
 
@@ -111,7 +111,7 @@ progress tracking.
 - [ ] Execute export and verification Prefect tasks on CPU OpenTofu modules,
       proving GPU instances are torn down before CPU provisioning.
 - [ ] Store FP32 exports under `/models/onnx/{experiment}/fp32/` with metadata
-      capturing opset, optimiser flags, and git references.
+      capturing opset, optimizer flags, and git references.
 
 ### 3.2. Enforce parity verification gates
 
@@ -122,13 +122,13 @@ progress tracking.
 - [ ] Schedule nightly parity spot-checks on retained models to detect drift in
       dependencies (CUDA, ONNX Runtime) before release branches freeze.
 
-### 3.3. Deliver quantisation and benchmarking workflow
+### 3.3. Deliver quantization and benchmarking workflow
 
-- [ ] Codify static INT8 quantisation scripts with calibration dataset reuse
+- [ ] Codify static INT8 quantization scripts with calibration dataset reuse
       and artefact write-back to `/models/onnx/{experiment}/int8/`.
 - [ ] Measure INT8 throughput versus FP32 on target CPU instances, documenting
       improvements and acceptable accuracy deltas (<1% MAE loss).
-- [ ] Integrate quantised-model verification into CI with `atol=1e-2`
+- [ ] Integrate quantized-model verification into CI with `atol=1e-2`
       tolerances, including latency smoke tests under realistic batch sizes.
 
 ## 4. Integrate with the Rust inference surface
@@ -151,7 +151,7 @@ progress tracking.
 - [ ] Benchmark end-to-end inference latencies (p50, p95) under representative
       traffic, ensuring INT8 pipelines meet sub-50 ms targets on t4g.large.
 
-### 4.3. Operationalise release governance
+### 4.3. Operationalize release governance
 
 - [ ] Define promotion criteria linking training experiment success, parity
       checks, and Rust e2e tests before artefacts become GA releases.

docs/opentofu-hcl-syntax-guide.md

@@ -19,7 +19,7 @@ infrastructure. OpenTofu, as an Infrastructure as Code (IaC) tool, reads this
 declarative configuration and takes on the responsibility of figuring out how
 to achieve that state.
 
-This approach is centered on the concept of state reconciliation. OpenTofu
+This approach is centred on the concept of state reconciliation. OpenTofu
 maintains a state file (by default, `terraform.tfstate`) that records the
 real-world resources it manages. When a configuration is applied, OpenTofu
 compares the desired state defined in the HCL files with the current state
@@ -620,7 +620,7 @@ provider's API.2 A data source is declared using a
 
 `data` block.
 
-#### Syntax and Behavior
+#### Syntax and Behaviour
 
 The syntax is `data "<PROVIDER>_<TYPE>" "<NAME>" {... }`.25
 
@@ -760,7 +760,7 @@ of the resource or module.5
 
 - **Syntax**: `count = <WHOLE_NUMBER>`
 
-- **Behavior**: It is best suited for creating multiple copies of a resource
+- **Behaviour**: It is best suited for creating multiple copies of a resource
   that are identical or vary only in ways that can be derived from a simple
   numeric index.
 
@@ -822,7 +822,7 @@ in the collection.29
 
 - **Syntax**: `for_each = <MAP_OR_SET_OF_STRINGS>`
 
-- **Behavior**: It creates a more stable association between the configuration
+- **Behaviour**: It creates a more stable association between the configuration
   and the real-world resource. Each instance is tracked by the map key or set
   value, not by a transient numeric index.
 
@@ -1000,7 +1000,7 @@ of operations for creating, updating, and destroying resources.
   dependencies is implicitly. When one resource's argument references an
   attribute of another resource (e.g., `subnet_id = aws_vpc.main.id`), OpenTofu
   automatically infers that the VPC must be created before the subnet. It
-  analyzes all such references to build the dependency graph.19
+  analyses all such references to build the dependency graph.19
 
 - **Explicit Dependencies with** `depends_on`: In some rare cases, a dependency
   exists that cannot be inferred from expression references. This typically