docs: track what the three-way review turned into, and gate CSS

[oratis]

Wrap-up for the batch that came out of #216.

Progress section, not a rewrite

docs/THREE_WAY_REVIEW.md §4 stays as the 2026-08-02 snapshot it says it is. A new §7 records what landed (F1–F7, F9–F12, F14 + the desktop palette), what was deliberately left (F13 — DeepSeek has no vision backend), and what is still open (F8 structured-item resume, F15 protocol thread/list/fork/archive, reasoning still dropped in the desktop, Linux bwrap not empirically re-tested).

Letting an assessment quietly reshape itself around what got built is exactly the drift it complains about.

Five things the implementation found that the review missed

Recorded in the same section, because they say something about where to look next:

• deepcode --version printed 0.1.0 on every release — it reads core's VERSION, which the release workflow never stamped
• the macOS sandbox denied reads of the workspace itself — (deny default) with no cwd rule, while Linux bound cwd read-write
• the desktop event envelope was overwritten by its own payload — review_action carries kind: 'apply', which clobbered kind: 'event'; invisible because there were no consumers
• skills/agents didn't de-duplicate by name — same-named skills described twice to the model, same-named agents resolved by scan order
• format:check didn't cover .css

The CSS gate

css joins the prettier globs. index.css was already clean, so this is a gate change with no reformatting — verified by appending an unclosed rule and watching format:check fail, then reverting.

🤖 Generated with Claude Code