docs: close out the 2026-08 push — roadmap table, memory, next steps

[oratis]

Summary

Docs-only close-out. Fifteen PRs (#6–#20) merged; main is green at server 656 / client 82 / 5 Playwright across five CI jobs — baselines were 377 / 0 / 0, and CI had been red since May.

• MASTER_PLAN §1 — roadmap statuses updated to the closing state, each row keeping its audit-time judgement in the notes. B3, C1–C5, D2, D5 complete; D1 and Hunter marked needs a human decision (ad accounts, paid-API budget) rather than left as "not started".
• MASTER_PLAN §2b — the surprise-defects table gains wave 2's three, including D3 was marked complete but had never worked and cache hits were being billed.
• MASTER_PLAN §4 — replaced "next up" with what's genuinely ready to pick up, ordered by value, and recorded what each wave-2 card did not cover (the usage ledger's blind spots; why SSE phases are coarse).
• memory.md §6 — rewritten: half the historical known-bugs list is closed, the rest restated alongside the new ones, and the footer now says plainly that none of this is deployed yet.

Three failure modes recorded in memory.md, because they generalize

1. Silent misconfiguration beats crashes. A nonexistent CSS token (ErrorCard rendering white-on-white), a nonexistent class (className="input" → unstyled native controls), a wrong function contract (llm.embed always returning null). None throw; all just quietly render or return the wrong thing. "Looks like an empty state" should be treated as a bug lead.
2. Testing only against fresh databases isn't testing production. The session index would have stopped the server booting against an existing DB while 563 tests stayed green.
3. Nullable columns are invisible data loss. workspace_id stayed nullable, so any INSERT that forgot it succeeded and then vanished from every scoped read.

Test plan

• Docs only, no code changes

🤖 Generated with Claude Code