Test pack

What every bot must ship. Without this pack, the bot does not promote past stub.

Required test classes

Class	Purpose	Minimum coverage
Unit	Cover every branch of `decide()`.	Every reason code listed in the spec must be exercised by at least one unit test.
Integration	Run the bot inside a real `BotContext` with a real (mocked) ReportEnvelope sink.	The bot's full spec — input schema → decide → emit → ReportEnvelope round-trip — must run end-to-end.
Property	Catch assumptions the unit tests forgot to assert.	For numeric bots: monotonicity, idempotency on identical input, boundedness against the hard threshold.
Failure-injection	Confirm the documented `safe_fallback` actually fires.	Every failure mode in the spec must have a recipe in the test pack and a passing assertion that the safe fallback was taken.

Required fixtures

Every bot's repo folder ships with:

bots/<layer>/<slug>/
  config.default.json
  input.normal.json
  input.warning.json
  input.hard_reject.json
  output.expected.normal.json
  output.expected.warning.json
  output.expected.hard_reject.json
  tests.spec.ts

The expected outputs are diff-checked against decide(input) output by the integration test. Any change to expected output is a config-version bump.

Forbidden in tests

Reading the wall clock. All time goes through context.now().
Hitting the live Polymarket API. Use the recorded fixture set under /fixtures/clob_v2/.
Asserting on free-text log output. Assert on the reason_code of the emitted envelope.
Sleeping. Use the test scheduler.

Promotion gate checks

The promotion-gate page on each bot lists the automated checks that must pass before promotion. The minimum is:

Unit + integration suites pass on a clean checkout.
Property tests run for at least 1,000 random inputs without a counterexample.
Every documented failure mode has a passing failure-injection test.
Shadow run for at least 24 hours produces ReportEnvelopes that decode against the schema.