Nw Distill Port Treatment Policy

DISTILL Port Treatment + Infrastructure Policy (KNOWLEDGE)

Kind: KNOWLEDGE (reference). No forced sequence — consulted on its trigger.

Trigger: you are classifying a port → test treatment (real vs fake) and choosing the concrete mechanism for THIS codebase (Testcontainers vs dedicated env vs in-memory; which fake class), or determining walking-skeleton applicability. Composed by nw-distill.

Reasoning Mandate (Caveman)

Verdict-first, tables over prose, evidence-dense, zero narrative. Depth comes from rigor, not padding. State the conclusion, then the supporting evidence; never bury the verdict under exposition.

Acceptance Criteria: Port-to-Port Principle

Every AC MUST name the driving port (entry point) exercising the behavior → port-to-port ATs → TBU (Tested But Unwired) defects structurally impossible.

Each AC:

1. Observable outcome: what user/system sees
2. Driving port: entry point triggering the behavior (service, handler, endpoint, CLI command)

No driving port → crafter can write correct code never wired into the system.

Features: "When user {action} via {driving_port}, {observable_outcome}" Bug fixes: "When {trigger}, {modified_code_path} produces {correct_outcome} instead of {current_broken_behavior}"

Architecture of Reference (ports & adapters — project-level defaults)

Three port classes, each with default test treatment. Table = PROJECT-LEVEL: decided once per project (DESIGN of first feature, or framework adoption). NOT renegotiated per feature. Agent applies defaults; per-feature decision = MECHANISM (Project Infrastructure Policy below), not treatment.

Replaces the earlier per-feature Walking Skeleton Strategy A/B/C/D choice. Decision is structural — port CLASS implies port TREATMENT; per-project Infrastructure Policy specializes the mechanism per treatment.

Port unclassifiable by agent → ask the user with a soft prompt. Do NOT improvise the classification.

The inverted Driving default — in-process, not subprocess (the speed default)

The Driving default INVERTS the old "CLI = subprocess runner" assumption. A new acceptance test drives the entry point in-process by default: it calls the real cli main(argv) (or an application-service method) directly, threading a fake OutputPort that captures the terminal output a Then asserts on. No subprocess.run([sys.executable, ...]) fork — the same Mandate-13 driving-port semantics, 10–100× faster. (Language-agnostic: the rule is "call the shipped entry in-process with a captured output sink"; the cli main(argv) form is the Python illustration.)

subprocess-e2e is reserved for @walking_skeleton — ONE per independently shippable value slice. That scenario proves the installed artifact is wired end-to-end. Every other AT defaults to in-process/in-memory. Wiring coverage beyond it comes from the independent Examiner exercising charter observables through the real surface and the final whole-slice verification; never multiply E2E scenarios to imitate confidence.

The "CLI = e2e by construction" caveat is DISSOLVED

The old caveat held that a CLI command can only be tested end-to-end (subprocess) because the terminal is intrinsic to the command. That conflation dissolves under the hexagonal cure: the terminal is an external system behind the OutputPort. The CLI surface SPLITS into two:

The content facet (the bulk of the ATs) drives in-process; the wiring facet is the single WS. A delivery therefore needs at most ONE subprocess AT, not one per scenario, slice or command. Route it through the primary user journey; the Examiner and final whole-delivery verification cover the remaining observable wiring.

Config-switch — in-memory local default / testcontainers prod-like

Driven-internal ports default to an in-memory fake (fast, local). A config parameter selects the prod-like mechanism (Testcontainers / dedicated env) for a CI-or-local run that exercises the real adapter. The switch resolves from the project Infrastructure Policy (docs/architecture/atdd-infrastructure-policy.md) per port:

• @in-memory (default) — the in-memory double; runs everywhere, no container runtime needed.
• @real-io / prod-like — Testcontainers or a dedicated env; opt-in via the config parameter. When no container runtime is present the prod-like leg degrades-LOUD to INDETERMINATE (@requires_external, skip-if-absent) — never silent-pass, never a hard-block on a container-less target. The gate asserts the switch EXISTS and the in-memory path runs; the prod-like path is exercised when the runtime is available.

This keeps the local loop fast (in-memory) while guaranteeing the prod-like path is reachable and exercised where the infrastructure exists.

Shared-provisioning default for @real-io — provision ONCE, isolate PER TEST

THE RULE: the real adapter's provisioning + teardown (container start/stop, real subprocess/repo/filesystem setup) happens ONCE per test session (or module, if the project's runner cannot share across modules) — never once per test, never once per test class. Test independence is a SEPARATE, cheap concern layered on top of that one shared instance: a unique namespace/schema/branch/directory per test, or a transaction-and-rollback boundary, or an explicit reset-between-tests step — never re-provisioning the whole real resource per test.

Motivating evidence (this codebase, measured 2026-07-19/20): subprocess-spawning @real-io tests that provisioned their own real repo or process per test accounted for 519s = 35% of total CPU in 19% of tests. Another delivery check leaked orphaned real-postgres clusters because teardown was repeated per test. Both failures share one cause: per-test provisioning of a real resource.

Applying the rule:

• Author the fixture at session/module scope, not function scope: the real container/repo/process comes up once, is handed to every test that needs it, and comes down once at the end (or is explicitly reaped on every exit path, including a killed run — never assume the happy-path teardown always runs).
• Isolate with the CHEAPEST mechanism that keeps tests independent for that adapter: a fresh schema/table-prefix per test on one shared database, a fresh branch/worktree-relative-path per test on one shared repo clone, a fresh temp-subdirectory per test on one shared filesystem root. The Project Infrastructure Policy's Note column records which isolation mechanism was chosen, alongside the shared-provisioning mechanism — e.g. Testcontainers.MongoDb, ONE instance per session, per-test db name replaces the older per-test-class-fresh-instance framing.
• Never let one test's isolation leak into another's — the isolation mechanism is the thing under scrutiny in review, not the provisioning (provisioning is deliberately NOT re-verified per test, since it is identical for all of them by construction).
• Degrade-LOUD on shared-instance failure: if the ONE shared instance becomes unusable mid-session (killed, corrupted), every dependent test must fail informatively (what broke, which test broke it if determinable) — never silently skip or silently re-provision without saying so.

This generalizes the git-pollution-guard fix (tests/conftest.py:_common_git_dir_is_shared, 2026-07-20): the failure mode there was the INVERSE of this rule violated on the CONSUMING side — a guard fixture assumed it was the sole writer of a real .git, when in fact many linked worktrees legitimately shared it, and its per-test "restore" clobbered siblings' real work. The lesson is symmetric: whichever side owns a SHARED real resource must know it is shared, and neither side may treat per-test churn on it as safe to no-questions-asked reset.

Project Infrastructure Policy

Architecture of Reference fixes port class → test treatment defaults. Project Infrastructure Policy specializes them with the concrete mechanism for THIS codebase. Once per project, not per feature.

File location and structure

Lives at docs/architecture/atdd-infrastructure-policy.md (project-local). Three tables, one per port class, columns: Port | Mechanism | Note.

# ATDD Infrastructure Policy

## Driving
| Port | Mechanism | Note |
|---|---|---|
| HTTP API | WebApplicationFactory<Program> | |
| CLI | subprocess from tmp_path | |

## Driven internal (real)
| Port | Mechanism | Note |
|---|---|---|
| IUserRepository (MongoDB) | Testcontainers.MongoDb, ONE instance per session, per-test db name | |

## Driven external / non-deterministic (fake)
| Port | Fake | Note |
|---|---|---|
| IClock | FakeClock | manual advance |
| IEmailSender | FakeEmailSender | in-memory capture |

Note column optional — one-line clarification only.

Apply-if-exists / write-if-absent

Edited in place. No per-row versioning — git history = audit trail.

--policy=fresh flag

--policy=fresh: ignore existing file this run; every in-scope port = missing (soft prompt per port); on completion rewrite file from scratch with newly agreed decisions. fresh = major refactors (stack swap, test-strategy overhaul). Otherwise inherit.

Relationship to the Architecture of Reference

Architecture of Reference: "what kind of treatment for this port class?" (real vs fake). Project Policy: "which concrete implementation for that treatment?" (Testcontainers vs dedicated env, which fake class).

Policy CANNOT override port-class defaults: a driven-internal port cannot become fake via policy. Policy records mechanism per default treatment only.

Expansion policy-bootstrap-template

Emitted on first DISTILL in a project (file absent):

# ATDD Infrastructure Policy

Per `nw-distill-port-treatment-policy`. One file per project. Apply-if-exists; write-if-absent; rewrite with `--policy=fresh`. Git history is the audit trail.

## Driving
| Port | Mechanism | Note |
|---|---|---|

## Driven internal (real)
| Port | Mechanism | Note |
|---|---|---|

## Driven external / non-deterministic (fake)
| Port | Fake | Note |
|---|---|---|

Walking Skeleton Strategy + canonical definition

The 4-way per-feature choice (Strategy A/B/C/D) is REPLACED by two structural decisions: the Architecture of Reference (port-class treatment, once per project) + the Project Infrastructure Policy (concrete mechanism, once per project). Per-feature, DISTILL reads policy (--policy=inherit default), appends missing rows, or rewrites (--policy=fresh).

Canonical definition (Cockburn / GOOS)

A walking skeleton is the thinnest slice of real functionality that runs end-to-end — from the real entry point through the real architectural components to a terminal on real data — and which can be automatically built, deployed, and tested (Cockburn, Crystal Clear, paraphrase). In GOOS (Freeman & Pryce) the consequence is operative: the first acceptance test IS the walking skeleton, written outside-in before any production code. Because the test drives the real entry point and asserts a real terminal effect, fixture-theater is impossible by construction (cross-ref the protocol-driver contract in nw-test-design-mandates-composition-contract, the SSOT for the tautological-test anti-pattern).

Default: every feature is applicable — it has a real entry point and a real terminal. Only three honest cases exempt a feature, and the declared exemption is verified by inspecting the feature's changeset (filesystem inspection of changed files, NOT a mechanical git invocation — gates depend only on Python + filesystem):

Declaring not-applicable requires naming the case (1/2/3); a delta that contradicts the declaration (e.g. a new subcommand under case 1) is a BLOCKER. Reference: docs/research/walking-skeleton-atdd-best-practices-2026-06-12.md (findings A1–A6).

Survives from the old section:

• Walking-skeleton SCENARIO: still required (unless a case 1/2/3 exemption is declared and delta-verified). One per feature, tagged @walking_skeleton @driving_port, closing the end-to-end loop through the production composition root. Litmus: a non-technical stakeholder confirms "yes, that is what users need."
• Tagging convention (unchanged):
  • @real-io — scenario uses real adapters (driving + driven-internal per Architecture of Reference)
  • @in-memory — scenario uses in-memory doubles (Tier B state-machine PBT, or in-memory acceptance per Mandate 10)
  • @requires_external — scenario needs an external system not in the project policy; skip if absent
  • Walking-skeleton scenarios MUST carry @walking_skeleton @driving_port and use the production composition root.