Add provider-neutral generation port and fake adapter

[chronoai-fkst]

Parent issue: #844
Parent PR: #846
Blocked reason: fix loop reached 12 rounds and was reconciled to blocked.

Smaller scope / alternative approach:
Governing practice: Ports-and-Adapters, contract-first fail-closed validation, and a single contract-owned canonicalization authority. This isolates the domain boundary after repeated drift in operation naming, result shape, validation ownership, and JSON array-preserving copies.

Smaller scope:

• Add TestCaseGenerationPort.generate(request) and the generation use case.
• Validate requests and successful candidate sets through the existing testing-design generation contract.
• Preserve the closed typed failure union and return detached snapshots from the fake adapter using contract canonicalization.

Non-goals: Codex CLI execution, Node process code, production runtime wiring, signed publication, or #795 handoff.

Acceptance criteria:

• Fake-adapter tests cover valid success, cancellation, every typed failure, malformed outcomes, foreign request_digest, prompt identity, and mutation isolation.
• No second canonical JSON or digest implementation is introduced.
• Valid arrays retain their JSON representation after detachment.
• Relevant checks and tests pass.

Non-goals:

• Do not repeat the same high-round fix path without reducing scope or changing approach.

Acceptance:

• The work is independently reviewable.
• The implementation can pass the normal intake, consensus, implementation, and review pipeline.

[wanghuan-520]

Human: Treat this decomposed child as the current highest-priority continuation of #844/#846 under #714 for the wanghuan-520 issue set. Run FKST Hosted intake, consensus, implementation, and review from the current #847 body using the reduced scope exactly as written. Preserve the non-goals: no Codex CLI execution, no Node process code, no production runtime wiring, no signed publication, and no #795 handoff in this child. Do not route through #652. Do not start #795 until this child and any remaining #844 decomposition children are resolved or definitively blocked.

fkst: reintake

⟦AI:FKST⟧

[chronoai-fkst]

⚠️ This issue carries an fkst work label but is not routed to any session.

Work issues are picked up only when exactly one assignee is set and that assignee is the creator of an active fkst session watching the label. Assign the matching session creator to route it; it will not be worked until then.

[chronoai-fkst]

👀 Picked up by fkst session mvp-integration.

A fkst pod is working this repo's fkst-dev-chronoai-fkst-cloud issues, including this one. The session posts its progress on this issue as it works, and the outcome will be a pull request (or, for issue-producing sessions, linked issues).

[chronoai-fkst]

github-devloop thinking: consensus started

[chronoai-fkst]

github-devloop decision: approve
Verdicts: teleology=abstain parsimony=approve fidelity=approve natural-ownership=approve proportional-containment=approve

teleology:
Ports-and-Adapters requires the application-owned port’s semantic result algebra to be defined before adapters and contract tests; otherwise the boundary cannot prevent provider-driven drift. The parsimony peer’s claim that enumeration is merely advisory does not move me: the issue says to “preserve the closed typed failure union” and test “every typed failure” plus cancellation, yet supplies no variants, fields, or authority. Current testing-design exposes only analyze (packages/testing-design/ports.lua:9, packages/testing-design/core.lua:39), while the generation contract closes receipt outcomes over complete, partial, rejected, budget-exhausted, and provider-error, excluding cancellation and any port-level result union (libraries/contract/testing_design_generation.lua:23). This is the root interface semantics, not a leaf implementation detail: developers could incompatibly treat cancellation, malformed adapter output, contract rejection, and provider failure as errors, tags, or receipt outcomes while all claiming acceptance. The reduced provider-neutral scope and recurrence response are otherwise sound, but the present authority failure makes development likely wrong. Enumerate the authoritative success/cancellation/failure algebra or cite its repository owner, then approve.

parsimony:
The peer attack does not defeat the root claim. Established Ports-and-Adapters practice places the result algebra with the application-owned port and validation with the existing contract. The issue advances that class-level boundary instead of repeating the Codex-specific fix loop, and its fake-adapter acceptance is code-verifiable. The peers correctly identify that “the closed typed failure union” is not enumerated: the current port exposes only analyze (packages/testing-design/ports.lua:9), and receipt outcomes are not a port-failure algebra (libraries/contract/testing_design_generation.lua:23). That is a necessary specification and PR-review checkpoint, not evidence that this bounded development proposal is false or likely to implement the wrong provider behavior. Define the success, cancellation, and failure variants before coding; do not silently reuse receipt outcomes or import CLI-specific failures. The contract remains the authority for request binding and candidate validation (libraries/contract/testing_design_generation.lua:216).

fidelity:
Ports-and-Adapters and contract-first validation govern this problem: testing-design should own the provider-neutral port, while the existing generation contract remains the authority for request and candidate-set validation (libraries/contract/testing_design_generation.lua:134, libraries/contract/testing_design_generation.lua:216). The peer claim moves me because this issue is a bounded development proposal, not a completed result-algebra specification. Its narrowed scope and fake-adapter acceptance are code-verifiable, and the absence of an enumerated “closed typed failure union” is a design decision to settle before implementation, not evidence that the issue is false or unsafe. This child also reframes the repeated provider-execution path by excluding CLI execution and production wiring. Advisory: enumerate success, cancellation, and each failure before coding; do not mistake receipt outcomes for port failures (libraries/contract/testing_design_generation.lua:23) or add a second canonicalization implementation when contract-owned copying already preserves array identity (libraries/contract/testing_design_generation.lua:105).

natural-ownership:
Ports-and-Adapters governs this problem: testing-design should own TestCaseGenerationPort.generate(request), while the generation contract owns validation, digests, and canonicalization. I previously abstained because “preserve the closed typed failure union” does not enumerate port failures. Parsimony’s claim moves me under the advancement standard: the omission requires grounding before coding, but the supplied issue and checkout do not establish that this narrowed proposal is false or unsafe. The issue replaces the repeated provider-execution path with a provider-neutral port and fake, excludes CLI and production wiring, and gives code-verifiable tests for malformed outcomes, request binding, prompt identity, and mutation isolation. The existing contract supplies request and candidate-set validation (libraries/contract/testing_design_generation.lua:134, libraries/contract/testing_design_generation.lua:216) and array-preserving canonical copying (libraries/contract/testing_design_generation.lua:105). Before implementation, enumerate the port’s success, cancellation, and failure variants; do not mistake the five receipt outcomes for that union (libraries/contract/testing_design_generation.lua:23, contracts/testing-design.v1.md:11).

proportional-containment:
The governing practice is Ports-and-Adapters: testing-design owns the generation port and use case; the existing contract owns validation and canonicalization. This child advances that boundary while expressly excluding CLI execution and process code, rather than repeating the #840/#842/#843 fix path. My Phase B objection relied on a parent-branch failure list that is not present in the pinned checkout; the current contract defines receipt outcomes, not a port-failure union (libraries/contract/testing_design_generation.lua:23; packages/testing-design/ports.lua:9). Thus the issue is sound, bounded, and code-verifiable, with no evidenced six-smell requiring abstention. Before coding, enumerate success, cancellation, and provider-neutral failures; do not inherit nonzero-exit or unavailable-binary as port meanings merely because an adapter may encounter them. Keep request and candidate validation contract-owned (libraries/contract/testing_design_generation.lua:134; libraries/contract/testing_design_generation.lua:216), and review how the fake obtains array-preserving detached copies, since canonical_copy is currently local (libraries/contract/testing_design_generation.lua:105).

⟦AI:FKST⟧