refactor(aidd-context): rework project-memory skill with capability model

[blafourcade]

🎯 What & why

Rework aidd-context:02-project-memory (renamed from 02-project-init). The old memory model was project-type / frontend-backend biased and the "init" name undersold a skill that also refreshes and reconciles. Now a capability model that fits any project shape and stays correct on re-runs.

🛠️ How it works

• Capability model (capability-signals.md): each capability is a definition plus open concrete evidence. A capability fires when a concrete repo fact matches the definition, never on an inferred domain, and every fire is surfaced for user confirmation. Why this way: a pure whitelist was brittle (missed custom auth), pure AI judgment was the old non-deterministic web-biased model. Definition + open evidence keeps determinism on concrete facts without the brittleness.
• Two false-fire edges fixed: auth fires on a custom oauth / credential / token module, not just named libraries; package fires only on an importable entry that is distinct from the CLI bin target (so a CLI no longer false-fires package).
• Actions stay tool-agnostic (Input/Output/Process/Test); action 04's test no longer contradicts its own flag step.
• Thin scaffold docs (README / GUIDELINES / CONTRIBUTING) point to the canonical framework docs instead of duplicating them, and the dead golden-principles / stack / workflows templates are gone.
• Memory templates regrouped one folder per capability, each macro-only with a macro Mermaid diagram.
• Hook (update_memory.js) now also refreshes a generated file index in memory/README.md and excludes that README from the always-loaded block.
• AGENTS.md template gains a communication section (answer-first, evidence over assertion, shortest decisive line, clarity carve-out for risky steps), borrowing the substance of the caveman skill, not its meme.

🧪 How to verify

• Regenerate the catalog clean: node scripts/summarize-markdown.js plugins/aidd-context/ /tmp/cat.md then diff against the committed CATALOG.md.
• Run the skill on a real project: in a repo with aidd-context enabled, ask in natural language to "bootstrap the project memory bank". It should detect capabilities from concrete signals, fill macro-only memory files, and sync the <aidd_project_memory> block + memory/README.md index.
• Validated this branch end to end via headless claude -p on the real cli project (clean init, reconcile, removed-capability) and a web-fullstack sandbox: detection was deterministic across clean and reconcile paths, auth fired definitionally, package was excluded on main===bin, and reconcile flagged a dropped capability without deleting its file.

⚠️ Heads-up

• CHANGELOG.md and aidd-context/CHANGELOG.md intentionally keep the project-init slug as shipped history; the rename lands as a new changelog entry on the next release.
• Full terse output stays opt-in (aidd-refine:03-condense); the AGENTS.md default is professional-concise, not caveman.

✅ I certify

• I DO CERTIFY I READ EACH LINE OF THE PULL REQUEST BECAUSE I AM A SOFTWARE ENGINEER, NOT A AI PUPPY.