Re-enable Tier-2 8-K event defenses (#9 + #10) in Phase 4

[dackclup]

Background

PR 3d shipped Tier-2 event defenses with 8-K Defenses #9 (Item 4.02 hard veto) and #10 (Item 4.01 auditor change) wired but disabled behind a module-level feature flag (compute.scoring.tier2._EIGHT_K_DEFENSES_ENABLED = False).

The deferral was forced by three workflow_dispatch timeouts (#12, #13, #14) on cold-cache runs:

• Run feat(phase-3d): Tier-2 going-concern + UI polish + fundamentals resilience #12: 45 min timeout. Cause: filing.text() routed through hybrid_section_detector (~5-10s × 502 stocks).
• Run feat(skills): scaffold 43 project-specific skills across all phases #13: 45 min timeout. Cause: filing.html() fix (commit 226840d) was insufficient — EightK.items / EightK.sections access ALSO triggered the detector via EightK.document cached_property. Fixed in commit 12ad7ff with regex-on-raw-HTML extraction (mirroring edgartools' Strategy 3 fallback).
• Run Re-enable Tier-2 8-K event defenses (#9 + #10) in Phase 4 #14: 1h+ stuck in fundamentals stage. Root cause: SEC EDGAR API throttling amplified by tenacity retry policy (60-90s/stuck-stock × ~40% failure rate). Fixed in PR 3d Part 1 (tightened retry + per-stock timeout + warning suppression) and Part 2 (per-stock latency histogram in Metadata).

After three timeouts the Tier-2 budget closed; we shipped going-concern-only.

Scope

Re-enable Defenses #9 + #10 with a re-architected fetch path that does not blow the workflow ceiling.

Implementation plan

1. Pre-cache layer (off-cycle nightly job):

   • New workflow compute-pre-cache.yml runs daily (or every 2 days) outside the weekly compute window.
   • Iterates the universe, fetches 8-K filing list per ticker, regex-extracts items + excerpts, writes to compute/cache/edgar_8k/<ticker>.json (existing 7-day TTL cache shape).
   • Weekly compute reads the warm cache; cold-fetch path becomes rare.

2. Flip the feature flag:

   # compute/scoring/tier2.py
   _EIGHT_K_DEFENSES_ENABLED = True

   Existing eight_k_enabled test fixture in tests/test_scoring/test_tier2.py becomes a no-op (or is removed).

3. Re-add non_reliance_filing to active veto layer:

   • compute/scoring/risk_overlay.py already accepts non_reliance_by_ticker: dict[str, bool]. The wiring stays valid — only the upstream input changes from "always empty dict" to "populated dict".
   • Update compute/main.py line ~580 to repopulate the dict from tier2_results.

4. Restore tier2_coverage_pct semantics:

   • With 8-K active, Tier2Result.fetch_succeeded reverts to "BOTH 10-K and 8-K succeeded".
   • Update tier2.py docstring and the relevant tests in section A.

5. Schema bump: 0.6.0-phase3d → 0.7.0-phase4-something. Active-veto count moves from 3 (PR 3d) back to 4 (PR 3e/Phase 4).

Testing requirements

• @pytest.mark.network integration test confirming a known 4.02 filing surfaces the veto.
• Schema snapshot guard catches the version bump.
• Workflow_dispatch cold-cache run completes in under 45 min with 8-K events active.

Acceptance criteria

• compute-pre-cache.yml workflow exists and runs successfully on schedule.
• _EIGHT_K_DEFENSES_ENABLED = True.
• non_reliance_filing appears in risk_flags for at least one veto-eligible ticker in production output.
• Workflow timeout returns to 45 min (revert commit 6b3c565).
• All section A tests pass without eight_k_enabled fixture.

References

• PR 3d perf hotfixes: 226840d (10-K filing.html()), 12ad7ff (8-K regex-on-HTML).
• SEC Form 8-K Items 4.01 / 4.02 (Reg S-K Item 304).
• Schroeder 2024 SSRN: ~50% of Item 4.02 filings precede formal restatement within 12 months.