Lane C: TDD wiring + zero-assembly default chain

[kyle-sexton]

Parent

Refs #2933. Source: topic pocock-shipping-breakdown (PR backfilled when opened).

What to build

Two related concerns:

1. TDD wiring. /tdd:principles exists but rarely gets invoked during implementation (known behavioral gap). Upstream's implement hard-wires "tdd at pre-agreed seams". Adjudicate candidates C9 (pre-agreed-seam gate: "no test is written at an unconfirmed seam" + the canned seam question), C10 (tautological-test anti-pattern: assertion recomputes expected the way the code does; expected values must come from an independent source of truth — complements our anti-gaming line), C11 (SDK-style per-operation mockable boundary functions). Decide where each lands: tdd:principles, testing:write, implementation:implement, and/or the review code lens.
2. Zero-assembly default chain. Upstream's implement proves a consumer gets a working implement→tdd→review→commit chain with no wiring. Verify our chain (implement → phase-verifier → verification:confirm → review:quality-gate → source-control:pull-request) composes out-of-the-box for a fresh consumer, and that every link is swappable/configurable rather than hard-coded. Document the default chain in one place a consumer can find.

Acceptance criteria

• C9/C10/C11 verdicts recorded in the SSOT; accepted ones landed
• A documented statement of the default execution chain and its override points
• TDD invocation path from implement is explicit (or its rejection recorded with reasons)

Blocked by

None — can start immediately.

[kyle-sexton]

Lane C adjudicated — PARTIAL

Design locked after adversarial validation: two fresh-context validators, rationale withheld, run over five proposed answers. All five were challenged. Two proposed ADOPT verdicts turned out to be already shipped, one proposed enforcement site structurally cannot host the rule it was given, and one disposal venue does not exist. The set below is the corrected one; it is recorded in the SSOT at docs/upstream/aihero-shipping-course.md under "Lane C (#2936)".

Landed in e846b691.

C2 + C9 — PARTIAL, relocated

C2 was routed here from Lane A as the same doctrine family, so both are adjudicated together.

Landed in plugins/planning/skills/plan/SKILL.md's existing Test strategy element — the one that already invokes /tdd:principles — and in its template placeholder. The plan names the public interfaces the tests will drive, each marked existing or newly introduced, with a preference for driving an existing interface over introducing one for testability alone. Naming them is what lets the Step 5 approval settle them, so implementation writes no test against a boundary the plan never named.

Two corrections against what was originally proposed:

1. Not phrased as "seams." seam is fleet-registered vocabulary — docs/conventions/seam-phrasing/ owns it for presence-gated cross-plugin references (37 uses as "tracker seam" alone), and architecture:improve enforces its own sense as controlled vocabulary that explicitly forbids substitution. Landing a third sense in planning would overload a registered term in the one plugin that has never used the word.
2. Not hosted by implementation:phase-verifier, the site originally proposed. That agent grades binary acceptance criteria against a final diff and is told "Everything you need arrives in your dispatch prompt … Refuse to guess either." There is no assumptions input, and the inline implement path does not dispatch it at all. "Boundaries stated before the first test" is a temporal-ordering claim no final-diff grader can observe — wiring it there manufactures INCONCLUSIVE-and-escalate loops.

Upstream's hard consent gate ("no test is written at an unconfirmed seam") is therefore softened deliberately, and the softening is the point: an unattended run cannot obtain confirmation, so a blocking gate there is a stall, not a safeguard. Instead, a boundary implementation picks that the plan never named is a deviation logged for PR-time review (DEVIATIONS.md beside PLAN.md in the contract slice) — this repo's existing unattended-assumption mechanism. The "ideal number of seams is one" numeric absolutism stays REJECTED per Lane A's folklore-figure posture.

C10 — ALREADY-PRESENT (prose)

Carried at plugins/tdd/skills/principles/reference/anti-patterns-khorikov.md:96,270 and, in checklist form, at plugins/testing/skills/write/context/write.md:78. No change.

Correction to a claim made while drafting this lane: the coverage is prose-only, not executable. testing/audit's cant-fail-scan.sh concedes in its own header that rule-recomputed-expectation "detects the decidable core — textually identical sides — not every recomputation shape," and a validator ran it over three canonical tautological tests for blocks that fired a block rule: 0. The canonical Khorikov shape — compute expected with the production algorithm, then assert — does not fire. The review-lens gap is genuinely open and is filed as #3046.

C11 — ALREADY-PRESENT + one-clause edit

test-doubles.md has carried ## SDK-Style Interfaces Over Generic Fetchers — the GOOD/BAD pair and "each mock returns one specific shape, no conditional logic in test setup" — since 85aa8066, predating the audit that proposed it. Adding it as "a small addition" would have produced a duplicate section.

The only genuine delta was the missing subordination clause: the section sits immediately below the five Khorikov mocking best practices but never stated a precedence relation to the first of them. That clause is now there — the shape rule never widens what gets mocked, and "mock only unmanaged dependencies" still decides whether a boundary is mocked at all.

Zero-assembly chain doc — REJECTED with reasons

The proposal was to write the default chain canonically into plugins/implementation/README.md and convert implement's scattered mentions into citations. Withdrawn on five defects, the decisive one being that the chain as drafted was factually wrong: the fresh-context phase verifier fires at every phase boundary (a loop) while confirm → quality-gate → pull-request fires once at completion, and that verifier is dispatched by implementation:implement-dispatch, not implement. Writing it down canonically would have laundered the error into an SSOT.

Also: READMEs are human-facing and graded on a different bar by docs hygiene; PLUGIN-PHILOSOPHY.md:474-476 routes a cross-plugin concern to a docs/conventions/ owner doc rather than a plugin README; converting the inline mentions to citations would strip the presence gates from the invocation sites, which the seam-phrasing convention places exactly there ("the gate belongs where the invocation is instructed"); and only 2 of 7 mentions actually overlap.

The acceptance criterion — "a documented statement of the default execution chain and its override points" — is served as it stands by implement/SKILL.md:186 + :199, with gates intact. The genuine find from this thread is filed as #3047: session-flow's pre-pr.md already owns an ordered pre-PR sequence, in a different order, declared unreorderable — a live SSOT conflict that a consolidation doc would have papered over rather than resolved.

The issue's own premise — UNVERIFIED, not STALE

"/tdd:principles exists but rarely gets invoked during implementation (known behavioral gap)" was first judged STALE because the invocation is wired at 7 sites. That inverts the claim: wiring is a property of text, invocation is runtime behavior, and seven wired sites plus rare invocation are jointly evidence that wiring is not the lever. Recorded as UNVERIFIED with wiring confirmed at 7 sites.

Measurement routes to the instrument that exists — claude-ops's SkillUse telemetry hook — and not to an evals item: plugins/evals/README.md:24-26 states "No command in this plugin executes model-graded evals," so an evals filing would produce JSON no runner executes. Filed as #3048, with the limit stated up front: that hook records no caller attribution, so implement→tdd co-occurrence is a proxy, not proof.

Board hygiene finding, larger than this lane

C1–C23 were defined only in the topic's memory slice (.work/, self-ignored, never committed). Every lane verdict on this board cited an artifact that dies with the working tree, and C11's entry was already stale — it listed as a gap something shipped since 85aa8066. All 23 definitions are now graduated into the SSOT as a durable candidate index. Any future proposal sourced from that C-list gets re-verified against the tree before it is believed.

Acceptance criteria

• C9/C10/C11 verdicts recorded in the SSOT; accepted ones landed — plus C2, routed here from Lane A
• A documented statement of the default execution chain and its override points — served by implement/SKILL.md:186 + :199 as they stand; the consolidation doc is rejected with reasons above, and the real ordering conflict is filed as SSOT conflict: implement's completion handoff prescribes a different pre-PR order than pre-pr.md, which declares its order unreorderable #3047
• TDD invocation path from implement is explicit (or its rejection recorded with reasons) — explicit at the "Before writing first test" chaining row; whether it fires is the open behavioral question, filed as Measure whether /tdd:principles actually gets invoked during implementation (SkillUse telemetry, not evals) #3048

Spin-offs: #3046, #3047, #3048.

---

Generated by Claude Code