feat(verify+auditor): OSAP factor-exposure proxy contract codification (closes #217 + #218)

[dackclup]

Summary

Closes #217 + #218 in one focused diff — both issues are the same root cause (Phase 4h OSAP factor-exposure proxy contract not codified anywhere) attacked from opposite sides of the audit surface.

The 2026-05-23 cron #3 surfaced a false-positive escalation chain: stock-detail-auditor read the universe-wide identical osap_signals dict (502 tickers × same dict) as cron-wide data corruption and woke incident-commander to a P1 that downgraded to P3 after walking the schema + frontend + blending pipeline. Root cause — the documented Phase 4h factor-exposure proxy contract (compute/features/osap_replicate.py:14-35, locked 2026-05-18) was nowhere in the agent prompts or the verify-helper. Both directions of drift (regression OR Phase 4i+ graduation) were silent.

What changed

File	Change
.claude/agents/stock-detail-auditor.md	Step 3 gains "Documented patterns — NOT broken_data" 3-row recognition table; new documented_proxy verdict; escalation row routing universe-wide scope-note-contract patterns to methodology-scientist (NOT incident-commander)
.claude/skills/verify-production-output/helper.py	New section_l_osap_proxy_invariant() — classifies OSAP surface into phase4h_proxy / blending_regression / graduated / schema_drift / unknown; wired into Section A-L main() runner
tests/test_verify_helper.py	+6 Section L cases (empty/skip + no-OSAP/skip + proxy-pass + regression-fail + graduated-warn + drift-fail + None-coverage-gap exclusion); 12 → 18 tests in this file
CLAUDE.md + AGENTS.md	Lockstep in-flight entry per §Conventions

Live data verification

Section L on the 2026-05-23 cron #3 output:

=== Section L — OSAP factor-exposure proxy invariant ===
  ✓ tickers with osap_signals: 502
  ✓ unique signal sets: 1
  ✓ distinct blended scores: 428
  ✓ mode: phase4h_proxy (1 signal set × 428 blended scores)

Section L passes; Summary 0 failures, 1 warnings (unchanged from pre-PR baseline).

Failure modes Section L catches

1. blending_regression — uniform signals AND uniform blended scores. Future code change that drops the per-ticker scalar multiplication in compute/scoring/osap_blend.py will FAIL the helper instead of passing silently.
2. graduated — varying signals (Phase 4i+ true per-stock replication). WARNS until the osap_replicate.py:14-35 scope-note is updated to declare the graduation explicitly — forces an intentional contract bump.
3. schema_drift — osap_signals key set differs across tickers (broken uniformity without graduating cleanly). FAIL.

Scope

No compute / schema / scoring / valuation / frontend code change — agent prompt + helper script + tests + lockstep docs only.

Test plan

• ruff check clean on touched files
• pytest tests/ -m "not network" → 1115 passed (7 skipped pyqlib/ipca env-missing, 22 deselected network)
• python .claude/skills/verify-production-output/helper.py on live 2026-05-23 cron feat(phase-1): universe + prices + momentum stub #3 data → Section L emits mode=phase4h_proxy (1 signal set × 428 blended scores); overall 0 failures, 1 warnings (unchanged baseline)
• stock-detail-auditor.md prompt update reviewed for clarity
• CLAUDE.md + AGENTS.md lockstep verified (both touched per §Conventions)
• CI green
• Spot-check on next cron feat(phase-2): SEC EDGAR fundamentals + per-stock detail pages #4 — Section L still passes; if a future cron sees mode=graduated without an intentional Phase 4i+ scope-note update, the WARN will surface it

---

Generated by Claude Code

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
quantrank	Building	Preview, Comment	May 23, 2026 9:35am