docs: add LEARNING.md capturing harness design reflections

[constk]

Problem

The harness encodes a lot of opinions (mechanical gates, meta-gates, layered boundaries, agent hooks) but the why behind those choices isn't captured anywhere a reader can find quickly. New contributors and users have to reverse-engineer the rationale from the gates themselves.

Proposed solution

Add LEARNING.md at the repo root, ~200–300 words, covering:

• What surprised the author while building the harness (where intuition was wrong).
• Which gate has caught the most real mistakes, and what kind.
• What would be done differently on a v2 (and why).

Keep it personal-voice and short — this isn't a design doc, it's a reflection.

Acceptance criteria

• LEARNING.md exists at repo root, ~200–300 words
• Covers: a surprise, a high-value gate, a thing-to-change
• Linked from README.md (e.g. under "About this project" or similar)

Priority rationale

Polish — not a blocker for release. Adds a humanising dimension that the rest of the docs (intentionally terse) don't carry.