feat: example 09 - best-time-to-implement (a task graph that schedules itself) - #7

What

A backlog modeled as a dependency graph that schedules itself. The "best time to implement" falls out of the graph structure, no flat to-do list. This is the raw, deterministic engine (state is data, every number hand-checkable); a later layer lets a model judge state from evidence at the assess seam, without touching the scheduling logic.

Four readouts, and the point is they disagree:

• READY - dependencies all done (startable now); else done or blocked (with the named unmet dep).
• Phi = value / cost (usefulFlowScore) - the ready set by what earns its cost first.
• Critical path - longest dependency chain by cost; sets the finish time.
• Sensitivity (rankSensitivity) - completing which task raises the most downstream available value/cost ("what unblocks the most").

The seed backlog includes the canonical case, domain-free: a task blocked on an unbuilt dependency ("move assets into object storage" waits on "build the storage adapter"), held as blocked until so it is not forgotten and surfaces when its blocker is done.

How

A cmg ModelGraph of three transforms, assess -> classify -> prioritize, each recording a signal (the run's trace shows how the plan was built). assess is the identity today; it is the seam for a model-call transform later. Reuses usefulFlowScore (core) + rankSensitivity (math). Built on main (the primitives from #6/#5).

Verify

pnpm --filter @composable-model-graph/example-09-best-time-to-implement start   # PASS
python3 python/examples/09-best-time-to-implement/main.py                       # PASS, byte-identical

TS and Python output is byte-identical (expected-output.txt); the self-check confirms every claim against the seed data. No keys, no network, no AI. Domain-free.

🤖 Generated with Claude Code