refactor(aidd-context): project-memory ships 23 templates and nothing says which a run reads

[blafourcade]

First proof of the doctrine written in #418. Slim aidd-context:02-project-memory, the heaviest skill outside the orchestrator, and check the contract survives a skill whose mass is in its assets.

Where the mass is

875 lines across 36 files. SKILL.md is 50; the rest is assets/templates/memory/*, thirty templates written to disk one at a time.

The question this issue must answer, and the contract must be able to express: does an action that writes one memory file read the other twenty-nine templates? If it does, the split rule is violated and the fix is structural. If it does not, the 875 lines on disk cost nothing per run, and the real target is the 50-line router.

That distinction is the whole point of measuring per path instead of per file.

Scope

• Measure the lines read on the shortest and on the longest path, before touching anything.
• Apply the contract from refactor(aidd-context): the skill template still teaches the format onboard abandoned #418; do not invent a local pattern.
• Preserve every current behavior.

Acceptance criteria

• The before/after lines-read-per-path is stated in the PR, shortest and longest path.
• The skill conforms to the contract from refactor(aidd-context): the skill template still teaches the format onboard abandoned #418 without needing an exception.
• No template is read by a path that does not write it.
• Nothing project-memory does today is lost.