Scale correctness for 1000-project workspaces + design-port wave 2

[Exelord]

Two things landed on this branch. The scale fix is the important one.

---

1. Scale: a 1000-project workspace stopped rendering empty pages and wrong ranks

Measured first. A probe seeding 1000 projects × 10 tasks (100k task_runs) through the real ingest wire showed the failure is correctness, not latency. /v1/projects?limit=500 returns a page, and three dashboard sources did fetch a page → .find() in it, so:

• every project past the page rendered a BLANK detail page (projectSummary found nothing),
• the ranking card claimed "vs 500 projects" when there were 1000, and ranked within the page,
• the Projects filter box could never reach a tail project (it filters the fetched rows).

Payloads were fine — 103 KB / 143 ms for 500 projects. The data was simply wrong.

Fixed server-side, the only place it can be right:

• listProjects(ws, {limit, search, projects}) — server-side ILIKE search + exact-name fetch (the point lookup).
• countProjects(ws) — the true denominator.
• rankProject(ws, project) — per-axis ranks via window functions over every project in one query, returning top-N per axis plus the named project with its true rank. Correct at any workspace size.
• /v1/projects answers {projects, total}; new /v1/projects/rank (allowlisted).
• The client-side ranker is replaced by a thin shaper; the Projects table now says "showing N of M projects" instead of implying the page is the workspace.

The guard caught a bug in the fix itself: = ANY($1) binds a JS array as a malformed array literal on this driver — IN ${sql(array)} is the form (the provenanceForHashes precedent). That would have 500'd the point-lookup route in production.

Pinned by a scale test seeding 620 projects (past the page): the tail project resolves by exact fetch and by search, ranks #1 by avg exec with total === 620, and a mid-pack project reports a true rank beyond the top-8 window. Wire pinned in the server e2e.

Still open (noted in the decision log): wiring server-side search into the table filter boxes (today the box filters the page and the notice tells the truth about it), and a point lookup for the flaky badge (getFlakiest(100).find() degrades for a task outside the top 100 — degraded, not broken).

---

2. Design-port wave 2

First wave shipped under the new visual guard (#166), which flagged exactly the three touched pages and left the other six pixel-identical.

• Callout — one banner primitive replacing five hand-rolled class strings across four views.
• deltaBar — signed deltas as a diverging bar around a shared zero with a flat band, so trivial noise renders neutral instead of the full danger red every non-zero delta used to get. Reviewing the capture caught a regression before it landed: the first cut mapped a task present in only one run to a zero delta, making "new" read as "no change".
• Ranking card — rankProjects had computed a hit-rate axis since July with no view rendering it. Third axis added; all three lists gained meters pinned to a 0..1 track.
• status FactField renders the shared badge instead of bare text.

---

Gates: oxfmt --check . clean · oxlint --type-aware --type-check 0 · cloud 550/0 · core from root 1286/0 · visual guard 10/10 (the ranking card renders identically off the new server source — free functional-equivalence confirmation). No schema/CACHE bump.

🤖 Generated with Claude Code

https://claude.ai/code/session_01RW7aso5j5CrBo7cjyET23D