scheduler: don't wait on input dependencies when cff.Predicate returns false

[Saijayavinoth]

Closes #104

Problem

When cff.Predicate returns false, the task body is skipped — but the scheduler still waits for every declared input dependency to finish first. An unrelated slow input imposes its full latency on the consumer, even though the predicate=false result is what decides the skip and the input value is never read. See linked issue for reproducer.

Change

Layer	What changes
API (scheduler)	PredicateJob{Run func(ctx)(bool,error); Dependencies []*ScheduledJob} + Scheduler.EnqueuePredicate. Re-exported as cff.PredicateJob.
Scheduler (3 sites in scheduler/scheduler.go)	Worker arm calls predicateRun; done-branch fast-dispatches consumers on !result && err==nil; enqueue handler does the same for consumers enqueued after the predicate completed (cached in ScheduledJob.predicateResult).
Codegen (predicate.go.tmpl)	Emits EnqueuePredicate(PredicateJob{}); Run sig changes from error to (bool, error). Consumer codegen byte-identical — its existing if !p<hash> { return nil } gate is untouched.

Safety

Fast-dispatch appears to race the slow producer's write to the consumer's input value. It does not:

• Skip gate precedes input read. Consumer wrapper is shaped if !p<hash> { return nil }; ... use(inputs...), so on predicate=false the function returns before any input is dereferenced.
• Slow producer (the dep we no longer wait on). When it eventually completes — success or error — the standard notification loop decrements consumer.remaining past zero (to −1). The if remaining == 0 push-to-ready check no longer matches, so the consumer is not re-dispatched. Producer errors still propagate to the flow's error path via the normal mechanism.
• Fast sibling producer errors. If a non-predicate input errors before fast-dispatch fires, the consumer is already marked invalid; fast-dispatch still pushes it to ready, and the worker's j.invalid check fires before any body code runs, surfacing the error as errJobInvalid — identical to pre-PR behavior under ContinueOnError.
• Late-enqueue race. Predicate may complete before the consumer is enqueued. Handled by caching predicateResult on ScheduledJob in the done branch; enqueue handler reads it and fast-dispatches inline.

Verification

• go test -race ./scheduler/... — passes, including new TestPredicateEarlyDispatch, TestPredicateTrueWaitsForDeps, TestPredicateError.

• make test — 21/22 internal/tests/ packages green. The one failure (TestPanicRecovered) is a pre-existing Go-version stack-format issue; reproduces with this change reverted.

• TestBlockingInputs_PredicateShortCircuits — asserts elapsed < slowDelay/10 for a 100ms slow dep.

• BenchmarkPredicateFalseWithSlowDep (internal/tests/benchmark/), 5ms slow dep, AMD EPYC 7B13:

  Variant	wall ns/op	consumer_ns/op
  baseline	5,226,110	5,219,513
  this PR	5,226,572	32,877

  Wall time bounded by Wait() draining the slow producer; consumer_ns/op is the metric of interest: ~159× reduction.

Out of scope (planned follow-up PR)

Fast-dispatch invalid consumers — same mechanism applied to consumers whose dep errored (j.invalid short-circuits without dereferencing inputs). Only meaningful under ContinueOnError; needs its own test coverage. Will open as a follow-up after this lands.

[codecov]

Codecov Report

❌ Patch coverage is 69.23077% with 8 lines in your changes missing coverage. Please review.
✅ Project coverage is 67.94%. Comparing base (4cdfe95) to head (9ab9f6e).

Files with missing lines	Patch %	Lines
scheduler/scheduler.go	69.23%	4 Missing and 4 partials ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #105   +/-   ##
=======================================
  Coverage   67.94%   67.94%           
=======================================
  Files          32       32           
  Lines        2059     2084   +25     
=======================================
+ Hits         1399     1416   +17     
- Misses        636      640    +4     
- Partials       24       28    +4     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Saijayavinoth]

@abhinav PTAL

[abhinav]

Apologies, @Saijayavinoth I don't work at Uber anymore. Someone from Uber can likely review.