fix(learning-readback): surface CONTEXT.md summaries instead of mangled dir slugs

[dd-solving]

Problem

In Releases/v5.0.0/.claude/hooks/lib/learning-readback.ts, loadFailurePatterns() reads each failure capture's CONTEXT.md into content and then never uses it — it injects the de-slugified directory name instead, hard-truncated at 70 characters with no word boundary. Every session start gets garbled lines like:

[2026-06-15] userfrustrated analysis incomplete demanding comprehensive audit of all sub

while the discarded CONTEXT.md contains a clean **Summary:** one-liner that FailureCapture.ts wrote at capture time. Additionally, if one event is captured twice (two dirs, different slugs, identical summaries), both lines are injected.

Fix

1. Prefer CONTEXT.md's **Summary:** line; fall back to the first line of ## What Happened; fall back to the de-slugged dir name (previous behavior).
2. Truncate at a word boundary (~110 chars) with an ellipsis instead of a hard 70-char mid-word cut.
3. Skip captures whose summary is already in the pattern list (dedupes double-fires).

Before / after (same capture)

before: [2026-06-15] userfrustrated analysis incomplete demanding comprehensive audit of all sub
after:  [2026-06-15] Frustrated — analysis incomplete, requesting comprehensive audit of all submissions

Behavior is unchanged when CONTEXT.md has no Summary/What-Happened sections — the slug fallback path is preserved.

Related (not in this PR)

loadLearningDigest() has the same discard pattern — it keeps an 80-char **Feedback:** stub and drops the Context paragraph. Happy to follow up with the same treatment if wanted.

🤖 Generated with Claude Code