feat(aidd-context): check existing rules for overlap or conflicts during rule generation

Start with plugins/aidd-context/skills/05-rule-generate/SKILL.md, then read 01-capture-rule.md, 02-write-rule.md, 03-validate.md, and references/tool-paths.md to understand the current flow and canonical rule surfaces. Add the pre-write assessment and approval gate without changing existing write or validation contracts; done means covered, overlap, conflict, needs review, and new cases identify relevant files and require an explicit user decision before writing.

Problem

aidd-context:05-rule-generate captures a rule intent, writes the rule for the selected target tools, then validates the generated files against the rule-authoring and tool-path contracts.

The workflow does not currently make comparison with the project's existing rules an explicit validation step before writing the candidate rule. As the rule corpus grows, a generated rule can duplicate an existing instruction, partially overlap it, contradict it, or introduce a different scope or priority for an already-covered concern.

This matters because repository-level AGENTS.md already requires checking whether an existing instruction, finding, or rule covers or contradicts a new one, then merging, removing, or clarifying it when necessary. The principle exists, but is not an explicit 05-rule-generate contract.

Scope

• Discover existing project rules relevant to the candidate before writing it through the canonical 05-rule-generate/references/tool-paths.md reference; do not add a tool-specific path list to this assessment.
• Assess the candidate policy, intent, and scope rather than comparing only its topic or title.
• Exclude expected cross-tool renderings of the same canonical policy from duplicate findings and filter rules whose scopes are disjoint from the candidate.
• Classify the candidate as new, covered, overlap, conflict, or needs review.
• Surface the existing rule files involved in every non-new result.
• Add an explicit pre-write assessment step between capture and write: 01 capture → 02 assess existing rules → 03 write → 04 validate.
• Stop before writing until the user has explicitly chosen how to handle the assessment result.
• Offer resolutions such as keeping the existing rule, merging, updating, clarifying scope or priority, creating anyway, or cancelling.
• Require explicit user approval before modifying an existing rule.
• Keep this as a guard inside 05-rule-generate, not a new skill.
• Preserve the existing write and validation contracts.

Acceptance criteria

• The workflow performs an explicit assessment step after capture and before writing.
• Rule sources are discovered through the canonical 05-rule-generate/references/tool-paths.md reference.
• Given a candidate whose policy is already expressed by an applicable existing rule, rule-generate identifies that rule before creating a duplicate.
• Given compatible but partially redundant instructions, the workflow reports overlap and identifies both rule sources.
• Given Always use pytest for unit tests and Use unittest for all unit tests, the workflow reports conflict before writing the candidate.
• Rules with disjoint scopes are excluded from overlap and conflict findings.
• Expected cross-tool renderings of the same canonical policy are not reported as duplicates.
• Each candidate is classified as new, covered, overlap, conflict, or needs review.
• Every non-new result identifies the involved rule files.
• The user makes an explicit decision before the candidate is written.
• The user can explicitly choose how to resolve an overlap or conflict.
• Existing rule files are never modified without explicit user approval.
• A candidate classified as new continues through the existing write and validation workflow.
• Existing structural, path, frontmatter, and per-tool validation behavior remains unchanged.

Prior art in this repo

• AGENTS.md#L27 already requires checking whether an existing instruction, finding, or rule covers or contradicts a new one before adding it.
• 05-rule-generate/SKILL.md#L8-L19 defines the current three-action flow: capture, write, validate; this issue adds the pre-write assessment step.
• tool-paths.md is the canonical source for supported rule surfaces and target paths.
• 01-capture-rule.md#L3-L28 captures topic, category, slug, scope, and write mode; it does not define semantic comparison with the existing rule corpus.
• 02-write-rule.md#L3-L26 builds and renders the rule for supported tools.
• 03-validate.md#L3-L22 validates written files against the authoring and tool-path contracts, but does not define a cross-rule consistency check.

I searched the repository's open and closed issues and pull requests for an equivalent semantic duplicate, overlap, or conflict proposal. Issue #788 concerns current Kilo/OpenCode output paths, not semantic consistency, so it is not a duplicate.

Out of scope

• Auditing the complete existing rule corpus independently of generating a new rule; that is #792.
• Automatically rewriting, merging, or deleting existing rules without user approval.
• Defining a universal precedence system between all AI tools.
• Changing existing rule file formats or target-tool mappings.