feat(squid): Brier score for predict-polymarket

[Tanya-atatakai]

Resolves BDMARK-2258

What

Adds Brier-score tracking to squids/predict-polymarket, mirroring the predict-omen implementation (see BRIER_SCORE.md), so the predict-economy Performance card can compute windowed Brier from the same query shape on both indexers.

• Bet.impliedProbability: BigInt! — per-token fill price, 1e18-scaled (|amount| * 1e18 / |shares|), set in handleOrderFill for v1 and v2 fills. Buys: the market-implied probability of the bet's outcome. Sells: closing price, excluded from aggregation. Zero = degenerate zero-share fill, skipped.
• DailyProfitStatistic.brierSum: BigInt! / brierCount: Int! — per buy, (p − actual)² / 1e18 with actual ∈ {1e18 won, 0 lost, 5e17 invalid}, credited on the resolution day inside processMarketResolution's existing per-participant bet loop. Windowed mean = Σ brierSum / Σ brierCount.
• New src/brier.ts with the pure primitives (computeImpliedProbability, brierContribution, actualForOutcome); invalid markets are keyed off winningIndex === -1.
• Migration 1787931352501-Data.js: three additive columns.
• Version bump to 0.3.0 (also syncs the stale 0.1.0 in package-lock.json).
• Docs: BRIER_SCORE.md polymarket section rewritten from "planning" to "shipped"; one line in root CLAUDE.md.

Omen's brierSumApplied/brierCountApplied reversal fields are deliberately absent: Polymarket resolutions are write-once (processMarketResolution returns early if a QuestionResolution exists), so there is no re-answer path.

Tests

• tests/brier.test.ts — fixed-point primitives (scale, sign handling, zero sentinel, no clamp, invalid = 0.5).
• tests/lifecycle.test.ts — new brier score block: win/loss sums on the resolution day, sells and zero-share fills excluded, invalid resolution, repeat-resolution idempotency; plus Brier assertions on the NegRisk path and impliedProbability on the existing buy test.
• 38/38 vitest tests pass; tsc clean.
• Migration verified against a fresh Postgres: migration:apply succeeds, subsequent migration:generate reports no schema diff.

Deployment

Requires a blue-green re-index from zero (historical bets need impliedProbability; running totals forbid partial replay). The current production stack stays untouched until cutover — procedure per the squid README.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AykA7kdEeusgyDs8dndpRx

[jmoreira-valory]

Review round 1 — ✅ Approve

What this does. Adds Brier-score tracking to the predict-polymarket squid so the predict-economy Performance card can compute windowed Brier from the same query on both indexers. Every order fill now stores its 1e18-scaled implied probability on the Bet, and at market resolution each buy's (p - actual)^2 contribution is credited to the agent's resolution-day DailyProfitStatistic (brierSum/brierCount), with invalid markets scored against 0.5. Requires a blue-green re-index from zero since historical bets lack the new column.

My previous blocker is resolved: the branch moved by one commit (b0d8626) that adds exactly the two missing allowlist entries — 1144965 (immutable 4.x, the range-sibling of the already-added 1144964) and 1145647 (node-tar DoS) — following the file's established same-GHSA-different-range twin pattern, and the previously-red Dependency audit (subgraphs/autonolas) job is now green. CI is fully green at HEAD (79/79 check runs, all 25 audit jobs passing) and the PR is mergeable.

The Brier implementation itself is unchanged since my last pass, and I had verified the load-bearing claims against source at HEAD rather than trusting the PR body:

• Omen parity is exact: PROBABILITY_SCALE = 1e18, (p - actual)^2 / 1e18, the zero-denominator sentinel, invalid → 5e17, and the brierSum/brierCount field names all match subgraphs/predict-omen/src/utils.ts, so the cross-indexer query-shape claim holds.
• Write-once resolution is real: processMarketResolution returns early when a QuestionResolution exists, and the per-participant settled guard gives second-layer idempotency — dropping Omen's brierSumApplied reversal fields is justified. Both the UMA and NegRisk paths funnel into it, and equal/short payouts correctly derive winningOutcome = -1 (invalid).
• The moved cache.set(DailyProfitStatistic, ...) (now after the bets loop) is safe: betsByParticipant's mid-loop flush only writes dirty entities, the in-memory mutations survive, and the deferred set lands in a later/end-of-batch flush. Participants are one-per-agent per market, so no cross-iteration aliasing.
• Sells store negative amounts and computeImpliedProbability absolutes them; sells and zero-share fills are excluded from aggregation by the isBuy && impliedProbability > 0n gate — all covered by tests.
• Tests are CI-enforced (the squid-tests job runs the vitest suite, green at HEAD), and the migration's NOT NULL-without-default columns are safe given the documented blue-green re-index from zero — they'd fail loudly if ever misapplied to a populated database, which is the right direction.

One out-of-scope aside carried over: the audit logs show ~40 allowlist entries past their review date — the backlog deserves its own sweep PR before the warnings drown real signal.

Findings — 1 · ⚪ 1 Nit

#	Sev	Location	Issue
F1	⚪ NIT	squids/predict-polymarket/schema.graphql:90	Zero sentinel also swallows zero-amount buys, not just zero-share fills

[jmoreira-valory]

Review round 1 — ✅ Approve

What this does. Adds Brier-score tracking to the predict-polymarket squid. Every order fill now records its 1e18-scaled implied probability on the Bet, and when a market resolves each buy's (p - actual)^2 contribution is credited to the agent's resolution-day DailyProfitStatistic, with invalid markets scored against 0.5. This gives the predict-economy Performance card the same windowed-Brier query shape on Polymarket as on Omen, at the cost of a blue-green re-index from zero.

Re-review — re-affirming my approval at HEAD. The branch moved by exactly one commit since my previous approve (b0d8626 → d8ffc6e), verified via the compare API: the applied F1 suggestion on the Bet.impliedProbability schema comment. One line, comment-only, no code, tests, migration, or allowlist changes. That push dismissed my earlier approval, so this review restores it.

• CI is fully green at d8ffc6e (30/30 check runs, including the dependency-audit and lockfile-validation jobs that were the subject of my earlier blocker), and @Adamantios has approved at the same commit.
• Nothing to re-verify in the implementation — src/brier.ts, src/logic.ts, src/handlers.ts, the additive migration, and both test files are byte-identical to the commit I already verified (Omen fixed-point parity, write-once resolution guard, deferred cache.set after the bets loop, sell / zero-fill exclusion via the isBuy && impliedProbability > 0n gate).
• Footgun sweep: the fleet-wide checklist (tx receipts, Safe execTransaction, TypedDict, CONTRACTS[chain], EIP-150/191, FSM sync, composed-app final states) does not apply to a TypeScript squid indexer. The subgraph-specific ones are moot here too: the migration is purely additive (no schema-drop / graft concern) and this is an SQD squid, not a graph-node file-data-source, so causality regions do not come into play.

The single NIT below is the leftover half of my F1: the suggestion block only covered schema.graphql, so the matching JSDoc in src/brier.ts still names only the zero-share case. Cosmetic; not worth another round on its own.

Findings — 1 · ⚪ 1 Nit

#	Sev	Location	Issue
F1	⚪ NIT	squids/predict-polymarket/src/brier.ts:9	computeImpliedProbability docstring still names only the zero-share case