Standardize label taxonomy and refinement label generation

[trsdn]

Problem

Issue labeling is still inconsistent across the system.

Examples that still show up:

• labels like research or enhancement that should map into the canonical type:* taxonomy
• labels like area:architecture showing up without clear rules about when area:* vs type:* should be used
• refinement-generated labels that do not always follow the expected naming conventions

This makes the backlog noisier and makes automation harder because the label vocabulary is not fully normalized.

Desired behavior

Standardize the canonical label taxonomy and make the refinement flow follow it consistently.

Acceptance criteria

• Define the canonical label families and intended usage, including at least type:*, priority:*, status:*, needs:*, and area:*.
• Document how legacy labels like research and enhancement should map into the canonical taxonomy.
• Update the refinement agent prompts/instructions/skills so generated labels follow the canonical naming consistently.
• Prevent new ad-hoc non-canonical labels from being introduced by the refinement flow.
• Add regression coverage or contract checks for label selection / normalization behavior.
• Existing repositories can still handle legacy labels safely during the migration window.