Manual-first when AI is down + the honest GitHub scorecard (/vs-github stops overclaiming) #5504

Why

Owner, 2026-08-19: "we shouldn't be solely relying on AI — if AI goes down, operators should be able to check and fix code manually" and "how do we compare against GitHub?"

Manual-first

The ship path (push → CI → gates → merge → deploy) was already AI-free by invariant. The operator/repair surfaces were AI-first; now every one has a working manual path:

• PR with conflicts — new src/lib/merge-conflicts.ts lists the conflicting files (git merge-tree --write-tree --name-only, fallback parser, never throws, no AI import); the PR page shows a Resolve conflicts manually panel with the files + exact commands, always — AI is the convenience.
• Repo health page — "Fix it yourself" block on every row (rule, file:line, change); "Fix this for me" disabled with the reason when AI is down; POST refuses honestly instead of 500.
• Issue page — "Re-run AI triage" disabled with by-hand guidance; honest info banner.
• docs/ops/MANUAL-FIRST.md runbook; ship-path-ai-independence.test.ts gains "operator surfaces offer a manual path".

vs GitHub

docs/VS-GITHUB-HONEST-2026-08-19.md — 19 dimensions with verdicts and what closes each gap. The public /vs-github page drops "one weekend built" / the fictional release-velocity chart, reads the Team price from the billing plans (it was a hardcoded fourth price), marks SSH as HTTPS+PAT on the hosted product, and says ahead/level/behind plainly.

Tests

Full suite: 4862 pass / 5 fail / 133 skip — the 5 are the pre-existing main failures. Typecheck clean.

🤖 Generated with Claude Code