Harden verified completion loop

[chriswu727]

What changed

• make complete user contracts authoritative before running redundant agent checks
• auto-submit verified work after successful commands, not only direct file edits
• reject empty test suites deterministically and return actionable discovery guidance
• preserve the latest local failure diagnostic while keeping older history bounded
• distinguish truncated previews from complete file writes and source files
• replace the lexical HTML benchmark oracle with parsed structure and row assertions
• require discoverable test artifacts plus visible external behavior assertions
• publish a fresh DeepSeek report and rewrite the README around Loop's core product idea

Why

The previous benchmark exposed three avoidable failure modes: a substring-based HTML
oracle rewarded removing <thead>, an exit-zero test command could execute zero tests,
and bounded observations hid the useful end of a traceback. Data-producing tasks also
spent extra turns re-reading artifacts or proposing checks already covered by the user
contract.

Impact

The final clean DeepSeek deepseek-chat run solves 12/12 cases with zero false
acceptances in 30 steps, 42,403 provider-reported tokens, and 65.795 seconds. Every case
is execution-verified, artifact-complete, integrity-valid, and replayable.

Validation

• make check — 428 API tests, 16 web tests, 14 desktop tests; lint, formatting,
  TypeScript, and mypy passed
• real-provider focused regressions for HTML, Python unittest discovery, data artifacts,
  and string utilities
• clean full 12-case DeepSeek run with matching manifest SHA-256
• Prettier and report/README consistency checks