Curation v2 step 2: Claude Opus 5 editor, the Brief, per-provider budgets

- AnthropicBackend (Messages API, cached system block, output_config.effort,
  server-side fallbacks, refusal surfaced as an error); Llms { bulk, editor }
  with editor_or_bulk(); PriceTable-based UsageMeter per provider.
- [anthropic], [editorial], deepseek.max_concurrent_requests and
  curation.max_article_count config; startup logs resolved providers.
- Budget day is the UTC date of started_at, preloaded from
  runs.provider_costs_json; finish_run writes provider_costs_json and
  config_json. Stage A batches run concurrently with per-batch budget checks.
- Editor prompt with one-line "why" per pick; no minimum lineup size;
  --max-articles is a ceiling; top-up branch deleted; why stored on picks and
  issue_articles.why and rendered in chapters and In this issue.
- Summaries on the editor client (3k-token input, concurrency 4, bulk then
  excerpt fallback); "The Brief" replaces From the Editor; section intros gone.
- Colophon carries per-provider costs and models.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A1rCLQeKBgnBo3oTgHuTMe

tests/fixtures/claude_brief.json

@@ -0,0 +1,3 @@
+{
+  "brief": "The lead, \"Migrating 40TB off Postgres\", is the rare migration write-up that keeps its failures in: two aborted cutovers, the rollback that took longer than the move, and a bill at the end. Read it first while the coffee is hot; it rewards attention and it is long.\n\nThe local desk answers with \"The MBTA slow-zone dataset\", which finally puts the T's own numbers into a shape a rider can argue with, and the charts do more persuading than a year of press releases. \"A failover story you'd argue with\" rounds out the engineering pages with a Postgres HA design that disagrees with the lead on almost every point, which is exactly why the two belong in the same issue. The issue is shorter than usual because a thin Friday is a good excuse to finish the long one properly rather than skim six."
+}

tests/fixtures/deepseek_front_page.json

@@ -1,8 +0,0 @@
-{
-  "from_the_editor": "Two of today's pieces are, underneath, the same story: what it costs to move data you no longer trust. The lead — a team hauling forty terabytes off Postgres, rollback plans and all — is the version with the invoices attached, and it earns the front page by refusing to tidy up its failures. Read it first, while the coffee is hot; it rewards attention and it is long.\n\nThe local desk supplies the counterpoint. Somebody has finally put the MBTA's slow-zone data into a shape a rider can argue with, and the charts do more persuading than a year of press releases. It is a short read and a satisfying one, and it pairs unreasonably well with the migration story: both are about institutions discovering what they actually have.\n\nThe rest of the issue is quieter than usual. That is not a complaint — a thin Friday is a good excuse to finish the long one properly rather than skimming six. If you only get through the lead today, you will not have missed much else.",
-  "section_intros": {
-    "Top Stories": "The day's most substantial piece: a full account of a forty-terabyte migration, with the failures left in. It is long, technical and unusually honest about what went wrong.",
-    "Boston & Local": "Transit data gets the treatment it deserves. A rider-built analysis of MBTA slow zones, with charts you can check yourself and a methodology section that holds up.",
-    "Niche Corner": "A section the model wrote an intro for even though nothing was placed in it today — a stray thread about tape-drive firmware and the people who still maintain it. The issue drops intros for sections that never ran."
-  }
-}

tests/fixtures/deepseek_lineup.json

@@ -1,10 +1,10 @@
 {
   "picks": [
-    { "id": 101, "section": "Top Stories", "position": 1, "lead_story": true },
-    { "id": 102, "section": "Tech & Engineering", "position": 1, "lead_story": false },
-    { "id": 103, "section": "Tech & Engineering", "position": 2, "lead_story": false },
-    { "id": 104, "section": "AI & Machine Learning", "position": 1, "lead_story": false },
-    { "id": 105, "section": "Boston & Local", "position": 1, "lead_story": false },
+    { "id": 101, "section": "Top Stories", "position": 1, "lead_story": true, "why": "The migration post-mortem with the invoices still attached" },
+    { "id": 102, "section": "Tech & Engineering", "position": 1, "lead_story": false, "why": "A failover story you'd argue with over coffee" },
+    { "id": 103, "section": "Tech & Engineering", "position": 2, "lead_story": false, "why": "Rare first-hand detail on a tool you use daily" },
+    { "id": 104, "section": "AI & Machine Learning", "position": 1, "lead_story": false, "why": "The one benchmark piece this week that shows its work" },
+    { "id": 105, "section": "Boston & Local", "position": 1, "lead_story": false, "why": "MBTA slow zones charted by a rider, not a press office" },
     { "id": 106, "section": "Culture & Essays", "position": 1 },
     { "section": "Niche Corner", "position": 2, "lead_story": false },
     "the model sometimes trails off like this"