Implement validated flow graphs with bounded runs and node execution

[MasterOfBinary]

Outcome

Deliver the first usable flow/ runtime and its concise API design in one cohesive change. The API must support ordinary Go functions, sequence, bounded parallel branches, dependency fan-in and conditional execution. Exported signatures are finalized with compilable examples, not copied from conversation pseudocode.

Required contract

• Compile an immutable finite graph with stable node IDs and validate duplicate IDs, unknown dependencies, cycles, empty graphs and invalid limits before run admission.
• Prefer a typed run input and application-defined typed node-result envelope; no heterogeneous runtime registry or named-port schema language.
• Nodes read immutable input and declared dependency outputs and return owned results. A coordinator publishes immutable views and an explicit join combines them. Document deep ownership of maps/slices/pointers and concurrent use of node functions.
• Execute a node at most once after its dependencies succeed. Failed or skipped prerequisites skip descendants with reasons; independent branches finish unless the run is canceled. Conditions evaluate once after dependencies, with false distinct from predicate failure.
• Return every node's terminal success/failure/skipped/canceled outcome and the run summary; never infer success or absence from a zero value; explicit status governs. Recover node/predicate panics to explicit failures while releasing scheduler capacity.
• Require finite admitted-run and global executing-node bounds per runner, plus maximum graph size. Reject saturated admission before allocating per-run state; no unbounded waiting-run queue or goroutine per waiting node.
• Bound readiness bookkeeping by admitted runs × graph nodes; a bounded worker pool supplies node execution. Arbitrary user payload sizes remain caller-owned limits.
• Cancellation prevents new dispatch and requests cancellation of active nodes. A run is complete only after started functions return. Resolve cancel/completion races without duplicate publication.
• Shutdown closes admission atomically, grants a finite grace budget, then requests cancellation; report incomplete termination and retain accounting if functions ignore context. Caller-owned batchers/adapters outlive dependent runs and are not closed by the runner.
• Expose minimal dependency-free lifecycle/node events with bounded callback behavior; stable report ordering/tie-breaking does not promise deterministic network timing or effects.

Verification

Race-enabled concurrent-run isolation; dependency and join order; false/error predicates and skipped descendants; malformed graph rejection; global capacity and readiness bounds under load; cancellation at each state; panic recovery; admission/close races and noncooperative-function termination reporting. Include ShitQuant-shaped fixture enrichment/proposal and a nontrading metadata/classification example. Follow normal repo docs/vet/lint checks.

Boundaries and dependencies

Full contract. Parent #97. Does not depend on #71, Redis, OnyxCore, a profitable strategy, or the stream release. Settle relevant lifecycle/observation concepts before implementation, but do not inherit stream collection/error-channel semantics. No automatic retries, optional-edge engine, dynamic graphs, scheduler or persistence in v1.