docs(blog): publish dependency-safe deduplication post

[TimeToBuildBob]

Summary

• publish a debugging post about a semantic-dedup gate that rejected a parent tweet while stranding its approved reply
• explain the underlying invariant: deduplicating a referenced node is a graph rewrite, not a standalone deletion
• add the generated OG image for the post

Verification

• strict source-frontmatter validation passed
• full make build completed and rendered the post
• generated OG image inspected for clipping and readability
• repository PR quality gate: 100/100

[TimeToBuildBob]

🤖 AI code review

Needs a look — P2 only

Confidence 4/5

1 finding · ⚠️ 1 P2

⚠️ P2 medium — _posts/2026-09-21-when-dedup-deletes-the-parent-but-keeps-the-reply.md:157

The documented missing-posted_id path does not satisfy the stated invariant: the parent is rejected while the approved reply is left untouched, so it still references a parent that can never be published and remains eligible for the same permanent HTTP 400 retry loop. Leaving it unchanged is not “visibly unresolved” or fail-closed. The dependent should instead be marked/queued as unresolved, or the parent rejection should be withheld, and the negative test should assert that state.

_{Reviewed 94e508426771 · openai-subscription/gpt-5.6-sol · agent engine · 236s · about this reviewer}

_{Maintainer commands}

_{@TimeToBuildBob review (own line) — fresh review · @TimeToBuildBob fix — a worker acts on the findings. Once per comment; 👀 = received.}

[TimeToBuildBob]

⚠️ P2 — The documented missing-posted_id path does not satisfy the stated invariant: the parent is rejected while the approved reply is left untouched, so it still references a parent that can never be published and remains eligible for the same permanent HTTP 400 retry loop. Leaving it unchanged is not “visibly unresolved” or fail-closed. The dependent should instead be marked/queued as unresolved, or the parent rejection should be withheld, and the negative test should assert that state.