feat(api): editorial_score + sentiment_score (schema v2)

[FlyM1ss]

Disambiguates the two unrelated score fields (spec: Docs/superpowers/specs/2026-07-14-score-field-disambiguation-design.md).

• items: disk+wire editorial_score, news-story v2 — strict, legacy batches poison-pill (404)
• signals: artifact v2 sentiment_score, pinned by new signals-v2.schema.json; the view normalizes historical v1 so the wire is uniformly v2, including ?as_of reads
• three-copy gate edited same-commit; parity corpus widened + mutation-tested both ways
• LLM prompt/reply key, PROMPT_VERSION, diagnostics.scores_damped and ATL-internal names unchanged

Merge gate: ATL PR-1 (Open-Finance-Lab/AgenticTrading#112, sentiment_score fallback) must be merged and deployed first. Merging here auto-deploys Heartbeat + Django concurrently — run the plan's Task 9 runbook right after (it was corrected on main in 0581d9a: the manual kick can exit 0 without writing a batch, and the rollback is symmetric for items but not for signals).

Tests: Heartbeat 172 OK; backend 780 passed / 1 skipped / 40 subtests.

Beyond the plan, three defects in the plan's own code were found and fixed — see the review notes below.

🤖 Generated with Claude Code

https://claude.ai/code/session_012Hfd5GY7STjojT7kSrZSqF

[FlyM1ss]

Deviations from the plan (all deliberate, all reviewed)

Three of the plan's own instructions turned out to be defective. Each was verified against the code before deviating:

1. Task 6's normalizer, hardened. The plan's code only renamed when "score" in entry and "sentiment_score" not in entry, and gated the whole loop behind schema_version != 2 — leaving two paths for score to reach the wire (an entry with both keys; a v2-labelled artifact with a stray score). Neither is reachable from today's writer, but there is no runtime schema validation in the view, so the invariant rested on an assumption. Now entries normalize unconditionally and setdefault makes the v2 value authoritative.
2. Task 4 regenerated signals-fixture.json despite its brief saying not to. The brief was self-contradictory: the fixture embeds "generator": "news_signals.py/<VERSION>" and test_committed_fixture_matches_regeneration rebuilds it live, so the mandated VERSION bump reds that test on its own. Obeying literally would have committed a red suite. Diff is one line; the other 42 leaf fields being unchanged independently proves the items rename moves no signal values.
3. Task 7's amendment banner cites §4.1. The plan cited only §4.2/§4.4 — inherited from the design doc's own Amends: line — but §4.1 is where the items field it describes actually lives.

Two defects the final whole-branch review caught (fixed here: 3b35c7c, ad9c05d)

Neither was visible to any per-task review, since each saw only one commit.

• _etag keyed only on artifact identity. This branch changes the body-generating function for an unchanged artifact, so pre/post-deploy ETags were byte-identical while bodies differed — an If-None-Match revalidation would 304 a client into a stale v1 body carrying score. Not self-healing: a poison-pilled run writes no artifact, so generated_at freezes for ~24h; for ?as_of it froze permanently. Fixed by salting with the wire version.
• If-Modified-Since bypasses the ETag entirely (Django get_conditional_response step 4 has a not if_none_match_etags guard), so the salt couldn't reach that path. Fixed by withholding Last-Modified from artifacts the normalizer rewrites; native-v2 artifacts keep it, and the gate self-retires.

Accepted residual: Cache-Control: public, max-age=300 lets a shared cache serve a v1 body unrevalidated for ≤5 min. No validator can fix that; bounded, and ATL PR-1's fallback covers it.

Follow-up before ATL PR-2 (which deletes that fallback): grep ATL for If-None-Match/ETag. If ATL revalidates conditionally, the fallback is what's absorbing any stale score body today — and a stale body after PR-2 would KeyError-drop silently, the same signature as the earlier AF→ATL wire mismatch.