Add explainable Signal strength scoring model (SIGNAL-SCORE-001)

[jamalfrnk]

Summary

Closes #37 (SIGNAL-SCORE-001). technical-analysis.ts already computes real indicators (EMA/MACD/RSI/ADX/Fisher/Keltner/ATR), but there was no unified, explainable 0-100 "Signal strength" score composing them into a documented, versioned model -- ruleAlignmentScore is a raw confirmation count, not that.

• signals/signalScore.ts -- pure, DB-free computeSignalScore(), same testability discipline as analytics/metrics.ts. Six weighted components (trend agreement 25, momentum agreement 20, trend-strength confirmation 15, volatility suitability 15, data freshness 15, indicator availability 10). Full weighting rationale in docs/product/signal-strength.md. Reuses technical-analysis.ts's own EMA-separation and ADX thresholds as normalization references rather than inventing new arbitrary constants.
• Conflict handling: trend/momentum disagreement (or an undeterminable trend) forces direction to NEUTRAL and the score to exactly 0 -- mirrors evaluateSignal()'s own choice to fire no signal at all in that case.
• Invalidation handling: RSI/Fisher pointing opposite the thesis, or price already breaking the Keltner band against it, caps the score at 15 without flipping direction -- the call hasn't reversed, only weakened.
• schemas/signalScore.ts -- versioned Zod contract (SCORE_MODEL_VERSION), mirrored on the client (no shared package exists between client/server in this repo, matching its existing convention).
• Wired into technical-analysis.ts's generateSignals(); persisted in a new nullable signals.signal_score jsonb column -- nullable because signals generated before this shipped never had a score computed, and backfilling a synthetic value for them would fabricate evidence that doesn't exist. Verified this decision was necessary: the shared dev DB had 2,431 pre-existing signal rows, and a NOT NULL migration would have been a real (blocked) data-loss operation.
• Client: SignalStrength.tsx replaces the old raw "Rule Alignment" bar with the new Signal strength score, a Strong/Moderate/Weak/None label (never color alone, per the mission's accessibility requirement), and a native <details> disclosure for the component breakdown -- no new UI dependency added.
• Never labeled "confidence"/"probability"/"expected return" anywhere, matching the ruleAlignmentScore precedent (GH audit finding F-5).

Verification

• 8 new unit tests cover every fixture category from the issue's test plan: full agreement (100), full disagreement (exactly 0), undeterminable trend (exactly 0), missing-indicator handling, stale-data handling, invalidation-condition capping, version stability, and the always-null candleInterval.
• Full suite: server 199/199 passing, client 33/33 passing. Both npm run build clean. Client lint/format clean.
• Verified against real market data end-to-end, not just unit tests: ran the real server against live Hyperliquid mainnet data, confirmed generated signals persist a correctly-computed score in Postgres -- including a real invalidation-condition case (RSI oversold correctly capping a live SHORT signal at 15/100) and a real non-invalidated case (component math cross-checked by hand against the persisted jsonb: (2*25+0*20+100*15+50*15+100*15+100*10)/100 = 48, matching exactly) -- and confirmed the API round-trips the jsonb value intact to the client-facing shape.
• Checked for CRLF pollution before committing (git diff --stat vs git diff -w --stat -- identical).

Test plan

• Independent reviewer reproduces the score computation against real generated signals (not just the diff), including the conflict-zero and invalidation-cap cases
• Confirm the nullable-column migration applies cleanly against a DB with pre-existing signal rows
• Confirm no regression in existing signal generation / API behavior

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a8a7a809d2

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[jamalfrnk]

Rework applied (independent review: FAIL -> addressed)

Independent review found a confirmed High regression: `signals.signal_score` is a nullable column by design (this DB alone had 6,728 of 6,731 rows genuinely null), but the client's `Signal.signalScore` type was non-nullable and `SignalStrength` dereferenced it directly -- any un-scored signal crashed with `Cannot read properties of null (reading 'totalScore')`, taking down the entire Signals page via the app's ErrorBoundary. Reviewer reproduced this directly with a throwaway test.

Fixed:

• `types.ts`: `Signal.signalScore` is now `SignalScore | null`, matching the real DB/API shape.
• `SignalStrength.tsx`: renders an explicit "Not available" state for a null score instead of crashing.
• New regression test (`SignalStrength.test.tsx`) covers both the null case and a real-score render.

Also addressed the two non-blocking Low findings:

• `signalScore.ts`: added an explicit zero-guard on `keltnerUpper`/`keltnerLower` in `volatilitySuitability`, matching the existing `ema200 !== 0` pattern.
• Two more test fixtures: momentum missing while trend is still determinable, and every indicator null simultaneously.

Full suite: server 201/201, client 35/35. Both builds clean, client lint/format clean.

[jamalfrnk]

Independent re-review: PASS

The confirmed High-severity crash (un-scored signal -> SignalStrength dereferencing null -> whole Signals page blanked via ErrorBoundary) is genuinely fixed, verified by reproduction: the reviewer ran the exact same <SignalStrength score={null} /> case the prior review used and confirmed it no longer throws, while the real-score path still renders correctly (no regression). Grepped the whole client for other unguarded signalScore consumers -- none found; SignalCard.tsx is the only other call site and passes through safely.

Both Low findings also confirmed genuinely addressed: the keltnerUpper/keltnerLower zero-guard is correctly placed, and the two new test fixtures (missing momentum, all-indicators-null) assert exactly what they claim, not just matching test names.

Full suite: server 201/201, client 35/35. Both builds clean. git diff between the reviewed commits is scoped to exactly the 5 files claimed -- no unrelated changes.

Ready for merge at your convenience.