Add read-only AL implementation guidance

[Jesper Schulz-Wedde (JesperSchulz)]

Scope and boundary

Adds a focused al-implementation-guidance capability above #155 for just-in-time Business Central AL authoring consultation. BCQuality remains read-only: it does not edit source, own workflow or consumed-guidance state, run an implementation/review-fix loop, compile, deploy, stage, commit, or publish. The consuming coding agent owns implementation, tests, retries, final review, and delivery.

Input and output contract

The action skill and host-native adapter accept an existing development-plan, readable repository, current implementation-diff, and current decision-context; optional consumed-guidance remains consumer-owned. Missing current diff or decision context returns not-applicable rather than inventing state.

A strict implementation-guidance-report preserves phase, decision, decision/evidence identifiers, applicability context, affected files/symbols/tokens, and consumer pins. Focused knowledge entries carry exact paths and optional checkout SHA, concrete used-for, faithful normative constraints, opened sample paths, validation considerations, suppressions, and unresolved applicability. Exact consumed path + decision-key + evidence-fingerprint matches are omitted deterministically; materially changed evidence can select the article again.

Retrieval and checkpoints

Retrieval is driven by current implementation evidence rather than broad plan domains. The skill recommends explicit, consumer-visible checkpoints for schema/data upgrade, public APIs/events/interfaces, permissions, external effects/job queues/HttpClient, telemetry/privacy, UI/page background tasks, tests, and bounded agent-initiated consultations. This is hybrid orchestration guidance, not hidden automatic behavior.

Evaluation evidence

Adds five production-shaped synthetic fixtures covering an expanded privacy-sensitive schema surface, exact consumed-guidance omission, a no-additional-guidance control, missing current context, and a read-only public-interface checkpoint. The shared scorer now validates both report kinds, exact deduplication, pins/context/reference integrity, and runner-owned pre/post target identity. The implementation regression is chained from the existing Windows/Ubuntu development-guidance CI entrypoint.

Local results:

• frontmatter validator: 0 errors/warnings
• review fixtures: 34 cases / 17 domains
• development-guidance fixtures: 5 passed
• development-guidance evaluator: 98 checks passed
• implementation-guidance fixtures: 5 passed
• implementation-guidance evaluator: 7 checks passed

Limitations

no-knowledge means no additional applicable constraints for the exact decision/evidence, not that code is safe, correct, complete, tested, or ready to release. Before/after evidence is not an OS write sandbox and cannot detect a transient reverted write. These fixtures do not run an agent, AL compiler, Business Central tests, deployment, or a production pilot. The documented baseline/plan-only/implementation-only/combined experiment is a recommendation, not a claimed result.

Dependency

Depends on #155 and intentionally targets jesperschulz-ingest-developer-knowledge, not main.

[Jesper Schulz-Wedde (JesperSchulz)]

Status: held pending #155 consumer evidence

Converting this PR to draft. The implementation is intentionally preserved, but it should not advance before #155 completes its explicit follow-up sequence:

1. agree the shared plan-guidance contract with a real consumer;
2. implement consumer uptake;
3. run the plan-only pilot and measure benefit;
4. examine which material late implementation/review findings remain.

Only then should we decide whether implementation-time guidance warrants a separate skill/report contract or belongs as a focused mode/input on the existing development-guidance interface.

Questions to answer from the pilot before reactivating this PR:

• Does plan guidance leave repeatable BC-specific authoring defects that implementation-time retrieval prevents?
• Is exact path + decision-key + evidence-fingerprint dedup useful in a real edit loop, or mostly inert as evidence changes?
• Can development and implementation guidance share one report contract with an optional focus block, avoiding parallel schemas and duplicated validation?
• What minimum consumer-owned state and checkpoint orchestration is practical?
• Does the extra retrieval improve first-pass results enough to justify added context, calls, and protocol complexity?

No consumer benefit is claimed by this draft. The current fixtures validate contract mechanics only; they do not establish that this capability improves generated AL beyond plan guidance plus independent final review.