Curation v2 step 7: cleanup, prune paths, implementation notes

Dead code and stale v1 comments removed (clippy -W dead_code clean, the
three world.rs warnings fixed), the Brief chapter's TOC title renamed from
"From the Editor", features prune now also sweeps article_assessments and
generate runs the sweep once after publishing, the example config is
tested key-for-key against Config::default(), README commands match
--help, and docs/plans/2026-08-15-implementation-notes.md records the
Anthropic and Voyage facts, the new tables, the budget-day rule and the
lock.

Implemented by a Claude agent from docs/plans/curation-v2-briefs/step7.md.

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

docs/plans/2026-08-15-implementation-notes.md

@@ -1,6 +1,8 @@
 # Implementation notes (shared brief for all implementation agents)
 
 Authoritative spec: `docs/plans/2026-08-15-the-daily-epub.md`. Read it fully before writing code.
+For curation (§3.5, §3.6 and §3.9 of that spec) the authority is now
+`docs/plans/2026-09-02-personalized-curation-v2.md`; see "Curation v2" below.
 This file records implementation-time decisions and verified external facts. Follow both.
 
 ## Verified external facts (2026-08-15)
@@ -68,6 +70,31 @@ This file records implementation-time decisions and verified external facts. Fol
   walks back up to `MAX_LOOKBACK_DAYS` and the section is datelined with the day it actually
   covers, not the masthead date.
 
+## Verified external facts (2026-09-02, curation v2)
+
+- **Anthropic Messages API** (verified 2026-09-02 against the bundled Claude API reference):
+  `POST https://api.anthropic.com/v1/messages` with headers `x-api-key`,
+  `anthropic-version: 2023-06-01`, `content-type: application/json` and
+  `anthropic-beta: server-side-fallback-2026-07-01`. Model id `claude-opus-5`; pricing
+  **$5.00 / M input, $25.00 / M output**, cache reads 0.1× input ($0.50/M), cache writes
+  1.25× ($6.25/M); the minimum cacheable prefix is 512 tokens. **No sampling parameters**
+  (`temperature`, `top_p`, `top_k` are a 400) and no `thinking` block — adaptive thinking is on
+  by default and depth is set with `output_config: {"effort": "high"}` (`low | medium | high |
+  xhigh | max`). The system prompt goes in `system: [{type: "text", text, cache_control:
+  {type: "ephemeral"}}]`; no assistant prefill, JSON is asked for in the prompt and parsed
+  tolerantly. `"fallbacks": "default"` (with the beta header) routes a request the safety
+  classifiers would refuse to a fallback model server-side; a response can still end with
+  `stop_reason: "refusal"` on HTTP 200, which the code treats as an error that degrades the
+  call to DeepSeek. Usage fields: `input_tokens` (uncached remainder),
+  `cache_creation_input_tokens`, `cache_read_input_tokens`, `output_tokens`. Timeout 300 s;
+  retry 429/5xx/network, never 400. Key only from `DAILY_EPUB_ANTHROPIC__API_KEY`.
+- **Voyage AI embeddings** (verified 2026-09-02): `POST https://api.voyageai.com/v1/embeddings`
+  with `Authorization: Bearer <key>`; body `{input: [...], model: "voyage-4-lite", input_type:
+  "document" | "query", truncation: true, output_dimension: 512, output_dtype: "float"}`. Up
+  to 1,000 inputs and 1M tokens per request, 32k tokens per input. Vectors are
+  unit-normalized, so dot product = cosine. **$0.02 / M tokens** after a 200M-token free
+  allocation. Key only from `DAILY_EPUB_VOYAGE__API_KEY`.
+
 ## Cross-cutting implementation decisions
 
 1. **sqlx usage**: use *runtime* queries (`sqlx::query(...).bind(...)`) and manual row mapping
@@ -80,12 +107,20 @@ This file records implementation-time decisions and verified external facts. Fol
    Pipeline stages are best-effort where the spec says so (social, XTC, world briefing, images).
 4. **HTTP**: one shared `reqwest::Client` (rustls, gzip, no cookies, 10s timeouts, UA
    `the-daily-epub/1.0 (personal rss digest; contact tyler@hallada.net)`), passed by clone.
-5. **LLM**: `async-openai` with custom base URL. All LLM calls go through `curate/llm.rs`
-   `LlmClient` which tracks token usage into a shared `UsageMeter` (input/cached/output tokens,
-   cost usd) and enforces `max_daily_usd`.
+5. **LLM**: a hand-rolled `reqwest` client, not `async-openai` (the published crate exposes
+   neither `Client` nor `CreateChatCompletionRequest` at the pinned version). Every LLM call
+   goes through `curate/llm.rs`: `LlmClient { system_prompt, model, meter, backend, retry }`
+   over the `ChatBackend` trait, with `DeepseekBackend` (OpenAI-compatible chat completions,
+   `response_format: json_object`) and `AnthropicBackend` (Messages API, facts above). The
+   pipeline holds `Llms { bulk, editor }`; `editor_or_bulk()` degrades to DeepSeek when the
+   Claude client is missing or its meter is tripped. One `UsageMeter` per provider
+   (DeepSeek, Anthropic, Voyage) with its own price table and `max_daily_usd`. The system
+   prompt is sent first and byte-identical within a run so both providers' prefix caches hit.
 6. **Testing**: unit tests inline per module; integration tests in `tests/` over fixture JSON in
-   `tests/fixtures/`. Never hit the network in tests. LLM stage mockable via `--skip-llm`
-   (prefilter order used for selection, feed excerpts as summaries).
+   `tests/fixtures/`. Never hit the network in tests: `MockBackend` (`ChatBackend`) and the
+   embedding mock (`EmbeddingBackend`) stand in for all three providers. `--skip-llm` makes
+   zero LLM calls (admission by cheap signals, `select_without_llm` by utility, excerpt
+   summaries); `--skip-embeddings` makes zero Voyage calls.
 7. **Style**: rustfmt defaults, `cargo clippy` clean-ish, no `unwrap()` outside tests, tracing
    spans per pipeline stage.
 8. **File ownership**: waves of agents work in parallel on disjoint files. Do not edit files
@@ -99,3 +134,40 @@ This file records implementation-time decisions and verified external facts. Fol
     `style-x4.css`). Askama 0.12+ configured via `askama.toml` if needed.
 12. **Determinism**: chapter ids `art-{entry_id}`, stable filenames, issue regeneration for the
     same date replaces prior rows/files (idempotent upsert everywhere).
+
+## Curation v2 (2026-09-02)
+
+The personalized ranker is specified in `docs/plans/2026-09-02-personalized-curation-v2.md`
+(§0 settled decisions, §3 target pipeline, §19 configuration, §21 the seven landed steps);
+`docs/plans/2026-09-02-curation-v2-progress.md` records per-step deviations. Facts an
+implementer needs that are easy to get wrong:
+
+- **Tables** (`migrations/0002_curation_v2.sql`, `0003_drop_scores.sql`; never edit
+  `0001_init.sql`): `rating_events` (append-only; the current verdict is the latest
+  `explicit` event), `article_embeddings` and `interest_embeddings` (f32 little-endian BLOBs,
+  `input_hash` = sha256 of the embedded text), `article_assessments` (`stage IN ('triage',
+  'deep')`, reused while `model` and `prompt_version` match and `assessed_at` is within
+  `assessment_reuse_days`; `--rescore` ignores the cache), `candidate_runs` (one row per
+  considered article per run, upserted with every column set on each stage transition),
+  `runs.config_json` / `runs.provider_costs_json`, `issue_articles.why`. `ratings`,
+  `feed_priors` and `scores` are dropped; `kv` keeps `ingest_watermark`, `taste_profile`,
+  `taste_profile_learned`, `profile_version`.
+- **Feedback**: `Vote` is `loved | good | down` (`NotForMe`); the HMAC message stays
+  `{issue_date}/{article_id}/{vote}`. `Vote::parse("up")` → `Loved` and `auth::verify_token`
+  still accepts tokens signed over the literal `up` segment because published issues carry
+  those links. Keep both.
+- **Budget day**: each provider's `UsageMeter` is preloaded with the spend of earlier runs on
+  the **UTC date of the run's `started_at`**, summed from `runs.provider_costs_json`
+  (`db::spend_for_date` by nominal issue date is gone). A tripped meter skips that provider's
+  remaining calls; the paper always publishes.
+- **Lock**: `src/lock.rs` takes `libc::flock(LOCK_EX | LOCK_NB)` on `<database_path>.lock`
+  for `generate`, `profile rebuild`, `features backfill` and `backfill-social`; a second
+  writer exits with "<command> is already running". `serve`, `explain`, `stats`, `ratings`,
+  `features prune` and `db migrate` never take it.
+- **Retention**: `telemetry::prune` removes `article_embeddings` of unrated, unpublished
+  articles older than `embedding_retention_days` (120) and `candidate_runs` rows plus
+  `article_assessments` older than `telemetry_retention_days` (180). `features prune` runs it
+  on demand; `generate` runs it once after publishing, best effort.
+- **Keys**: `DAILY_EPUB_ANTHROPIC__API_KEY` and `DAILY_EPUB_VOYAGE__API_KEY` map onto
+  `AnthropicConfig.api_key` / `VoyageConfig.api_key` through figment; the fields exist only
+  for that mapping and are never documented in TOML, logged, or stored.