fix(news): isolate the backtest signals projection

[FlyM1ss]

get_news_sentiment projected via a dict comprehension, so one off-spec ticker raised out of the whole projection and load_news_sentiment's fail-closed except emptied the sentiment slot for every ticker that step — a producer typo on MSFT silently deleting NVDA's news. The panel has dropped-and-carried-on since _representative_feed; this is the projection that never learned it.

Per-entry drop + the existing _alarm_if_all_dropped. Three symptoms, one cause — the adapter left contract-break reporting to its caller:

• Isolation. Blast radius drops from "every ticker" to "the broken one". Safe because the dict is already partial by nature (only tickers with signals appear), so a dropped ticker is the shape callers already handle.
• Layering. backtest_backend's except no longer has to explain _project_entry's KeyError from another module. That hint is now removed — it would go stale the moment the adapter's internals move, which is what this PR does.
• The second caller. get_news_sentiment has one production caller today. The fail-loud property fix(backtest): log news-sentiment failures #116 added lives in the caller's try/except, so caller Jifeng #2 (live paper-trading — anticipated in finsearch-news-signals-panel-design.md:115) silently regresses unless it remembers to re-implement it. Now there is no exception to forget to catch.

Isolation without the alarm would be a regression, not a fix. A rename makes every entry unreadable, so log-and-drop alone yields an empty slot plus a few warnings — indistinguishable from a quiet news day. That is how the 2026-07-14 outage ran for hours. Per _alarm_if_all_dropped's docstring, the alarm is what separates "one story is off-spec" from "the contract moved".

The alarm's raw arg is the post-filter set, never the wire's signals block. A signal for a ticker outside the universe is filtered by design, not dropped by drift; feeding the raw block would fire on every narrow-universe run and train the reader to ignore the one alarm that matters. test_universe_filter_alone_is_not_drift pins it.

_project_entry stays the single source of which fields are required — no duplicated key list, and KeyError's own text names the missing key, so the log reads missing 'sentiment_score' rather than a bare "malformed".

Escalation is log-only, deliberately. The panel escalates drift to a degraded badge; the backtest has no equivalent channel. RunManifest.news_sentiment_source is the candidate but is a provenance field that is unwired today (always None) — filing a follow-up rather than half-wiring it into a status flag here.

Mutation-tested (all three new guards fail when reverted; the alarm-deletion mutation was re-run after the first attempt produced a vacuous syntax error rather than a real failure). Suite green, delta is exactly +3 and nothing else moves (main 990 -> 993 passed, 2 skipped). Absolute counts vary with optional deps — discord installed or not swings the total, which is where an earlier 980/4 reading came from.

Review fix (ca36d12, docs only). _alarm_if_all_dropped's docstring claimed "the caller escalates a True to degraded" — true for the panel's call sites, false for the one this PR adds, which discards the return by design. Same hazard this PR removes from backtest_backend.py, one function over; #123 already quotes the line as if it were universal. Also NVDA-for-AAPL (AAPL is the one watchlist ticker the fixture gives no signals to) and the test docstring no longer re-narrates the history verbatim.

Re #120: independent (get_news_sentiment vs get_latest_panel_payload), and test-merged clean. Both widen _alarm_if_all_dropped to Union[list, dict]; this makes that change byte-identical to #120's so it merges either order. #120's "applied to ALL THREE projections" is left alone — it enumerates the panel's projections and stays accurate.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Pbs5wnQZhTi9CQ3pb14vDo

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agentic-trading-lab	Ready	Preview, Comment	Jul 15, 2026 3:52pm

[FlyM1ss]

Reviewed. One real finding, docs-only, fixed in ca36d12.

_alarm_if_all_dropped's docstring promised an escalation its new caller doesn't make — "the caller escalates a True to degraded" holds for the panel's call sites, not the backtest one added here, which discards the return by design (log-only pending #123). It's the hazard this PR removes from backtest_backend.py, one function over: a comment narrating another layer's behavior goes stale the moment that layer moves. #123 quoting the line as if it were universal is that drift already starting. The docstring now says what a True is worth is the caller's to decide, and names the backtest gap.

Plus two nits: NVDA-for-AAPL in the isolation docstring (AAPL is the one watchlist ticker the fixture gives no signals to — the one that couldn't be deleted this way), and the test docstring no longer re-narrates the history verbatim from _project_step_signals.

Verified rather than assumed, on head c95b7ae:

• Mutations — reverting isolation → comprehension, deleting the alarm call, and swapping considered → the raw signals block each turn exactly the intended test red, and only that one. Re-run after the docstring edits.
• fix(news): guard the panel signals projection #120 — git merge-tree clean both before and after ca36d12; the identical signature widening collapses.
• Suite — 993 passed / 2 skipped; delta vs main is exactly +3.
• The alarm actually lands. These loggers inherit root's WARNING (uvicorn's LOGGING_CONFIG has no root key), so error() escapes to stderr via lastResort. Worth stating because it's one level of margin — the same line as logger.info() would have been silently dead in prod.

Design call looks right: isolation without the alarm would have converted a loud failure into a quiet one, which is the 2026-07-14 shape. No new security surface — entries are dropped, never coerced, so the LLM context can only get smaller, never malformed.

[FlyM1ss]

/simplify pass — 4 review angles (reuse, simplification, efficiency, altitude). One fix applied, rest skipped as scoped-out or rejected on merits.

Fixed (516f05f, docstring-only). load_news_sentiment's docstring spelled out the adapter's drop-and-alarm mechanism in the sentence before warning that a comment here "goes stale the moment they move" — the rule and its violation two lines apart, and the same species of claim ca36d12 removed one layer down. Now points (see get_news_sentiment) per the file's existing idiom, so the rationale lives once.

Skipped — scope. _project_step_signals is the 4th hand-written loop+isolate+warn projector; reuse and altitude flagged it independently and both said don't fix here. The sharp version: the _alarm_if_all_dropped call is a separately-rememberable step, not forced by the primitive — which is how #119 happened. Fixing it rewrites three functions this PR never touches. Worth a pass if a 5th caller lands.

Rejected on merits. Unifying _project_step_signals with #120's _project_panel_signals: one is a pass-through filter, the other a transform. The panel can pre-check keys cheaply because it converts nothing; the backtest can't, since float(sig["published"]) only fails during the transform. Forced abstraction.

Confirmed correct, not changed. The escalation asymmetry (3 callers OR the bool, this one discards it) — #123 is real and states why deferring is right. Altitude added the clincher: api/v2/runs.py:367 writes the manifest once at run creation, before the field could be mutated per-step, so wiring it is genuinely #123-sized. Efficiency: no findings — the extra dict pass is ~microseconds/step against a cached fetch, ~0.1s across a multi-year backtest.

Suite 993 passed / 2 skipped, unchanged from this PR's own baseline.

Follow-up (docs, not touched here). This PR makes stale: finsearch-news-items.md:96 ("Drift escalates status to degraded" stated universally — now false for the backtest call site), :52 ("a rename of a field ATL doesn't read is invisible here" — false after #120 and this PR), and the Traceability table (103-114, no row for the new guard). Also finsearch-news-sentiment.md:141-153's failure-mode table has no row for the drop/ERROR behavior this PR completes.